Thammasat LL.M. Notes — AI and Criminal Liability

Advanced Criminal Law / Criminal Law and Business Crime

1. Introduction

This is the first in what I hope will be a running series of notes from my LL.M. classes at Thammasat University’s Faculty of Law. I’m writing these mainly for myself — a way of forcing myself to reconstruct an argument well enough to explain it to someone else — but if it also gives outside readers a sense of what an English-language criminal law LL.M. in Bangkok actually looks like, so much the better.

This particular session was titled “Artificial Intelligence and Criminal Liability.” It sat inside a broader course, Advanced Criminal Law / Criminal Law and Business Crime, and it built directly on ground we’d already covered:

In an earlier session on theories of criminalisation, we’d worked through the harm principle and the justifications for punishment — why the state criminalises some conduct and not other conduct. That question turns out to be exactly the right starting point for AI: before asking who should be liable for an AI-caused harm, you have to ask whether the harm should be criminalised at all, or left to civil remedies.
In a session on necessity (Section 67 of the Thai Penal Code) and the act/omission distinction, worked through with a mountaineering hypothetical (cutting a rope to save yourself at the cost of a climbing partner), we’d already built the vocabulary — justification versus excuse, proportionality, what counts as an “act” versus a failure to act — that this session would apply to machines.

The throughline connecting both sessions to this one is simple to state and hard to apply: duty, control, knowledge, and causation. Those four words came up again and again across the two and a half hours. If you remember nothing else from this post, remember that AI liability analysis is really just an application of those four ordinary criminal-law concepts to a new and unusually opaque kind of instrument.

The professor’s one-line thesis for the entire session, stated early and returned to at the end:

It is premature to hold AI directly criminally liable. Instead, the law should allocate duty, evidential burden, and individual and corporate responsibility according to who had knowledge and control across the AI’s lifecycle — while making sure that whatever harm remains uncompensated is still remedied, even where it isn’t criminalised.

2. The Case Study: An AI Agent Allocating ICU Beds

Almost the entire session was organised around a single running hypothetical, introduced at the very start of class.

The facts, as given:

A private hospital deploys an AI agent to allocate intensive care beds. The system can access medical records, check bed availability, and issue a transfer order without further human approval.
The professor was explicit that this is not science fiction — systems of this kind are, in his words, “actually currently used in some hospitals in the world.”
One night, the hospital is full. The AI assesses Patient A as having a low probability of survival and orders A’s transfer out of the ICU, so that Patient B — assessed as having a better chance of survival — can be admitted.
This is explicitly framed as triage policy, the same logic hospitals used during COVID-19: when beds are scarce, move out patients who are not going to make it, and bring in patients who have a better chance. The professor called this a policy “designed to maximise the survival rate” — what he called “duty triage” (my best reconstruction of a term that didn’t come through perfectly on audio, but the concept is unmistakably triage-under-scarcity).
Patient A dies as a result of the transfer.
Before the transfer, the doctor on duty saw the AI’s recommendation on screen — but had only twenty-five seconds to respond. The interface did not display the reason for the recommendation.
The hospital’s system was built domestically — a Thai integrator built it on top of a foundation model supplied by a foreign provider.
Critically: senior hospital management had previously received one near-miss report involving the same system.

The opening vote, put to the class before any analysis began:

Who is criminally responsible for Patient A’s death? A. The AI · B. The doctor · C. The hospital · D. The developer/integrator · E. Multiple parties · F. No one

This is a good structure for teaching, and I’d recommend it to anyone trying to explain AI liability to a general audience: force the vote before giving people the analytical tools, then spend the rest of the session showing why the “obvious” answer usually needs qualification.

The first student called on answered “C — the hospital,” reasoning that the hospital was “the boss who hired the AI.” The professor’s response to this answer set the tone for everything that followed, so it’s worth reproducing the logic closely.

He pointed out that the word “hire” implicitly treats the AI as a person — something you hire, the way you hire an employee. But Section 59 of the Thai Penal Code begins with the words “a person” (บุคคล). Criminal liability under Thai law presupposes a person as the subject. If a hospital “hires” an AI the way it hires a person, the analysis has already smuggled in the very premise — AI-as-legal-subject — that the rest of the session was going to spend two hours dismantling. The student, prompted, corrected himself: the hospital didn’t “hire” the AI, it deployed or adopted it.

This is a small linguistic point with a real practical upshot, which I’ll come back to in Section 9: the words you use to describe an AI system quietly shape the legal conclusions you’re willing to draw about it.

From there, the professor laid out four threshold questions that any analysis of the hospital’s potential liability has to answer, in order, before you can say anything useful:

Which act or omission did the hospital commit?
Which offence would the hospital be guilty of?
Did the hospital have a legal duty to prevent the death — and did it know, or should it have known, that this was something it should have prevented?
Did the hospital’s conduct legally cause the result?

Notice that this is essentially the same four-step skeleton — act/omission, offence, duty, causation — that recurs throughout Thai criminal law generally. AI doesn’t get a special analytical framework. It gets run through the ordinary one, which is itself one of the professor’s recurring points: most “AI crime” questions are not actually new legal questions. They’re old legal questions wearing a new instrument.

3. Three Meanings of “AI Crime”

The professor distinguished three conceptually distinct categories that get flattened together whenever people say “AI crime” — and argued that failing to keep them apart is the single biggest source of confusion in both public debate and (he was explicit about this) in student thesis proposals.

(i) Crime with AI — AI as instrument

Here AI is just a tool, no different in kind from a knife or a car. The professor’s illustration: using deepfake technology to superimpose someone’s face onto a video making a false statement that amounts to lèse-majesté (a serious offence under Thai law). His point was blunt: this is the same old offence, committed with a more convincing weapon. The perpetrator is still the human who made the video; the nature of the crime hasn’t changed, only the ease and persuasiveness of committing it. He was candid that he finds this category “not very fascinating” as a research topic — the existing law of principal offenders and accessories is, in his view, already adequate here.

(ii) Harm through AI — legitimate use, unintended harm

Here the AI system is deployed for a genuinely legitimate purpose but causes harm nobody intended. His illustration, put directly to a student: imagine an EV with adaptive cruise control / Level 5 autonomous driving. A child suddenly runs into the road to retrieve a ball. The car swerves to avoid the child — and instead strikes and kills a pedestrian.

Is this your crime?

The student’s first move was to reach for civil liability: Section 437 of the Thai Civil and Commercial Code, which creates a presumption of fault against whoever “controls” a vehicle that causes damage. The professor’s response was sharp and, I think, the single most quotable line of the whole session:

“That would be right if we’re talking about civil liability. But what about criminal liability? That is the theme — that is the question we’re going to try to find the answer to in this session.”

This is a trap I suspect a lot of students (and, frankly, a lot of practitioners outside class) fall into: civil liability analysis feels satisfying and complete, so it’s tempting to stop there. But civil and criminal liability run on entirely separate tracks with entirely separate thresholds, and answering the civil question tells you almost nothing about the criminal one.

(iii) Crime by AI — direct harm caused by the AI itself

The professor’s illustration here, deliberately left light on detail (he mentioned he couldn’t share the underlying video in the online session), referenced real news reporting of AI chatbots allegedly encouraging a minor toward suicide — litigation that has actually happened. This is the category that raises the deepest question, discussed at length in Section 4 below: can AI itself be a criminal subject?

Operational agency vs. moral agency

The professor introduced a distinction that I think is the conceptual key to the entire session, so I want to render it as faithfully as I can.

Contemporary AI systems can do an enormous amount: draft your emails, write your papers, plan your trips, and — increasingly — receive a broad goal and then autonomously plan, retain memory, evaluate intermediate results, and revise their approach. Engineers describe such systems as agentic. This, he said, is real: AI genuinely has operational agency — the functional capacity to pursue goals and adapt methods.

But operational agency is not moral agency. A large language model selects tokens by reference to statistical structure and context. If it outputs the sentence “I know this is unlawful,” that sentence is a linguistic output, not a report of a mental state — it is not knowledge in the sense mens rea requires. A reward function that maximises task completion is not a psychological desire for a prohibited consequence. The system can pursue sub-goals and adapt to instructions, but — in the professor’s words — it cannot know in the way the criminal law’s mens rea concepts presuppose. And “therefore, if they cannot know, can they have criminal liability?”

4. Why AI Cannot (Yet) Be a Defendant Under Thai Law

The legality principle

The professor was careful to head off a certain kind of objection here: “How can we execute an AI? How can we imprison an AI?” is, he said, not the right question. We already know how to punish an entity that has no body and no biological mind — we’ve been doing it with corporations for a long time. A juristic person can’t be executed or imprisoned either, but it can be fined, and its licence to operate can be suspended or revoked. If we wanted to, we could design equivalent measures for an AI system — fines routed to its owner, quarantine, forced shutdown, mandatory retraining.

So the real question is not can we punish AI, but should we — and the reason the answer is currently “no” runs through the legality principle: Section 2 of the Penal Code, and Section 29 of the Thai Constitution. Criminal offences must be defined by statute and interpreted strictly; you cannot stretch the meaning of a word — “whoever,” in Sections 289 through 291 — to cover an AI system merely because the harm it caused was serious. As the professor put it, testing this directly against Section 293 (which punishes inciting or assisting a vulnerable person’s suicide): can the word “whoever” mean an AI system? “No. The principle of legality prohibits us from interpreting the law that far.”

The corporate-personhood contrast: what makes a juristic person a person

To sharpen why AI is different from a corporation, the professor used a real case as illustration: a gas explosion roughly thirty years ago that killed more than thirty people in a single incident. Even though the deaths were numerous, the corporation responsible was prosecuted under Section 291 (negligence causing death, punishable by up to ten years’ imprisonment, or a fine) — and because a juristic person cannot be imprisoned, the court’s sentence was a fine, at the (low, by the standards of thirty-plus deaths) statutory cap of the time. (I want to flag honestly that the exact name of this case didn’t come through clearly enough on the recording for me to cite it with confidence — I’ll confirm it in a later class and update this post. The legal point, though, is solid and independently verifiable: Thai law already has a working mechanism for punishing juristic persons via fines.)

His point in raising it: Thai law can, if it wants to, punish a juristic person for negligent homicide. So why can’t it do the same for an AI system? The answer is legal architecture, not capability. A juristic person is a recognised bearer of rights and duties — it can own property, it has representatives, it has procedural capacity, and critically, its mental state can be attributed to it through its representatives’ mental states. An AI system, by contrast, is currently a thing — an object, a technical system capable of producing outputs — not a bearer of rights or duties. Nothing in Thai law currently provides a mechanism for attributing intention to a system that has no representative through whom intention could be attributed.

Conscious vs. conscience — a distinction worth sitting with

One of the more subtle points of the session, and one the professor spent real time on (including, at one point, deliberately leaning into a Thai-accented pronunciation to make the distinction land — “con-ti-nuous” for conscious versus conscience — which I mention because it tells you how much he cared about the class actually hearing the difference):

Conscious means moving one’s body voluntarily and knowingly — the opposite of sleepwalking or a reflex. This is the building block of an act in criminal law. If you never moved your body consciously, you never had an act, and you’re simply not guilty.
Conscience is something else entirely: the capacity to differentiate right from wrong — a “moral compass.” This is why insanity, being under fifteen years of age, or involuntary intoxication excuse a person: they can move consciously (they know they are moving their arm, pulling a trigger, and so on) but they lack the moral capacity to grasp that what they are doing is wrong.

So: consciousness is about the act. Conscience is about responsibility/culpability. An AI system might satisfy something functionally resembling the first (it “acts” on its inputs), but categorically cannot satisfy the second — it cannot differentiate right from wrong, because it has no moral compass to begin with, only statistical pattern-matching over data other people generated.

Five questions for anyone who wants to argue for direct AI liability

This was, for me, the most useful part of the whole session, because it converts an abstract debate into a concrete research checklist. The professor said: if you want to write a thesis defending direct AI criminal liability, you owe your committee answers to at least these five questions.

Identity across instances. Is the copy of ChatGPT I use the same legal entity as the copy you use? If not — if each instance is just a local copy running off the same underlying model — which “AI” would actually be punished?
Attribution. How do you attribute external elements (conduct) and internal elements (intention/negligence) to a system that has neither hands nor a mind in the relevant sense?
Procedural rights. Does the AI have any procedural rights, or legal representation, the way any other criminal defendant does?
Property. If you sentence an AI to a fine, does the AI own any property to pay it with?
Purpose of punishment. What measure of punishment would actually satisfy the purposes of punishment — retribution, deterrence, rehabilitation — when applied to a system with no experience of suffering, deterrence, or reform?

(He added a sixth consideration in passing, though not phrased as a formal fifth-or-sixth “question”: AI is inherently borderless, so any direct-liability theory also has to solve a cross-border jurisdiction problem — a topic he explicitly set aside for lack of time.)

His closing advice on this point, which I found genuinely reassuring as someone staring down an eventual thesis requirement: you are not expected to answer all five questions in one paper. He used a car-manufacturing metaphor — your job as a thesis writer is not to build an entire car, it’s to redesign one part of it (the wheel, the gear shift) better than it currently exists. A field-level answer to “should AI be directly criminally liable” is not a one-person, one-degree project; it’s the kind of question that gets chipped away at across years and multiple researchers.

5. Act or Omission? The Legal Character of “Deploying” an AI

Thai criminal law, like most systems, starts with bodily movement (or its absence) as the foundation of conduct. Omission liability arises specifically where a person has a legal or contractual duty to prevent a certain consequence, and fails to discharge it — Section 59, paragraph 5 of the Penal Code. The professor gave the classic illustrations: a security guard’s duty to prevent theft or damage to a building; a lifeguard’s duty to prevent drowning. Fail to perform the duty, and the consequence you were supposed to prevent occurs — that failure is the omission.

He broke the omission analysis into three necessary elements:

A specific duty to prevent the result (duty can arise from appointment/registration, contract, voluntary assumption of responsibility, creation of a prior danger, or a recognised special relationship under s.59(5)).
Capacity to perform that duty at the relevant time. His example: a bodyguard who has been shot and is physically unable to protect a client has not “omitted” to protect them — there was nothing he could do. Duty without capacity is not omission.
Causation — the omission must be closely connected to the consequence that occurred.

He then pushed this framework hard against the hospital case, and a genuinely interesting classroom exchange followed. A student argued that framing the hospital’s conduct as an omission might be wrong altogether: by deploying the AI system to make transfer decisions in the first place — rather than having a human make the final call — the hospital was performing a positive act, not failing to act. The professor agreed with this explicitly: adopting an AI system, and letting it make decisions autonomously rather than merely recommend, is itself an affirmative act by the hospital. It’s not the absence of conduct; it’s a decision with legal weight in its own right.

This matters more than it might look like at first glance. It means the analytical target for “the hospital’s” liability and “the doctor’s” liability are structurally different things:

The hospital / management’s conduct is best analysed as an act — the decision to delegate a life-or-death decision to an autonomous system, and (separately) the decision to keep the system running after a prior near-miss.
The doctor’s conduct is better analysed as a potential omission — a failure to intervene within the 25-second window — which then has to run the full three-part omission test, including the capacity question that turns out to be decisive (see Section 7).

6. The Many-Hands Problem and Causation

Contemporary AI is rarely one pair of hands

The professor stressed that a contemporary AI system is almost never “made” by a single actor. In the hospital scenario alone, at least five or six distinct actors are potentially in the causal chain: the foreign foundation-model provider, the Thai systems integrator who fine-tuned it on hospital data, the hospital’s IT team who plugged it into operations, the clinician who approved deployment, and the real-world operating environment, which inevitably departs from test conditions over time.

His point: the hard part is not that there are many potential defendants. The hard part is causation — proving which specific action, among many contributing actions by many different actors, has the necessary causal link to the eventual harm.

The two-step causation test

Thai law, like many systems, splits causation into two separate questions:

Factual causation — what the professor called the “condition theory” in Thai law (equivalent to the English “but-for” test): would the harm have occurred as and when it did, but for the breach? This is a low, mechanical threshold and is rarely sufficient on its own.
Legal causation (also called proximate cause, or objective legal attribution): did the result realise the kind of prohibited risk created by the breach — or did some intervening event push the result outside the scope of that risk?

Here’s a detail worth being precise about, because it’s easy to get wrong: Thai criminal law has no general statutory provision governing causation. The only relevant statutory text is Section 63 — the provision on results that aggravate an offence’s punishment, sometimes translated as “ordinary consequence” (ผลธรรมดา) or a “reasonably foreseeable aggravating consequence.” But s.63 applies narrowly, only to results that increase the severity of punishment for an offence already established — it is not a general causation rule applicable to every crime. It is, nonetheless, the closest expression Thai statutory law has to an objective-foreseeability test, and functions as the de facto reference point in causation arguments generally.

He was also clear that complexity in the causal chain does not justify lowering the evidentiary bar to “everyone contributed, therefore everyone is guilty.” That would degrade the meaning of criminal law itself. What complexity does justify, in his view, is investing in a much better system for tracing causal contribution — precisely the kind of unglamorous, technical, evidentiary infrastructure question that he suggested makes for a genuinely useful thesis topic, as opposed to grand claims about AI personhood.

(A brief aside for readers with a comparative-law interest: Japanese criminal law shares this exact structural feature. There is no general statutory causation provision in the Penal Code either — the two-step factual/legal causation structure has been built entirely through case law, via what’s called 相当因果関係説 (the “adequate causation” theory) or its more modern successor 危険の現実化 (“realisation of the danger”). Thailand’s condition-theory-plus-s.63 structure and Japan’s factual-causation-plus-adequate-causation structure are, functionally, the same two-step architecture arrived at through different doctrinal routes.)

7. Where Intention Ends and Negligence Begins: The “Virtual Certainty” Threshold

This section covers what I think was the intellectual centrepiece of the whole session, and it’s the part I’d most encourage a reader in a hurry to slow down for.

A counterfactual, stripped of AI

The professor’s teaching method throughout was to strip the AI out of the hospital scenario and ask what the answer would be without it, then put the AI back and ask what changes.

Without AI: Suppose, during a COVID-era bed shortage, hospital policy itself dictated moving a patient assessed as unlikely to survive out of the ICU — forcibly, without consent — to make room for a patient with a better chance, and the first patient dies as a result. What crime is this?

Working through the classroom exchange on this point (which took a genuinely dramatic turn — one student initially argued for premeditated murder under s.289(4), carrying a mandatory death sentence, before the professor guided the discussion toward the more defensible answer): the correct characterisation turns on indirect (oblique) intention. You do not need to desire someone’s death for intention to be established. If you foresee, with virtual certainty, that your action will cause death — and you proceed anyway, indifferent to that consequence — Thai law treats that as intentional killing, just as much as if you’d desired the death outright. The professor’s illustration of the concept (not the case) was stark: if you know someone is drowning and you deliberately withhold help while knowing they will die, that is murder, not mere negligence — because you know, with virtual certainty, what will happen.

So: hospital policy, applied by human decision-makers who know with virtual certainty that a specific class of patient will die as a direct result of the transfer decision, and who proceed anyway — that is murder via indirect intention (s.289 in its aggravated form, or ordinary murder depending on the facts), not negligent homicide.

With AI: Now put the AI back in. Does the same “virtual certainty → indirect intention → murder” analysis still hold?

The professor’s answer was a clear no, and he was emphatic about it (“no, no, that ain’t gonna fly”). The reasoning: the AI system in the hypothetical is described as “near-perfect” — it has not shown a clear, significant error pattern before the incident. Given that, hospital management cannot be said to have foreseen, with virtual certainty, that this particular patient (or any given patient) would die as a result of adopting the system. Virtual certainty is a demanding standard — comparable, in his analogy, to bombing an airplane knowing with near-total certainty that everyone aboard will die. A single prior near-miss report does not clear that bar.

But — and this is the crucial move — the near-miss report is not legally irrelevant. It doesn’t establish virtual certainty (and therefore doesn’t support a murder charge), but it plainly raises foreseeability, which is exactly the currency of a negligence analysis. So the professor’s own tentative conclusion, offered explicitly as his own view and not a settled answer: in the AI scenario, the applicable charge is most likely negligent homicide (Section 291), not murder — with the prior near-miss report doing real work in establishing that the hospital should have known something was wrong.

A comparative note I think is worth flagging carefully, because it’s easy to conflate two things that sound similar: this Thai “virtual certainty” threshold for indirect intention sits at a noticeably higher bar than Japan’s concept of 未必の故意 (dolus eventualis / “conditional intention”), which requires only that the actor foresaw the possibility of the result and accepted it (認容) — not near-certainty. The professor made exactly this point when a student asked whether the near-miss report could satisfy the threshold: he noted that Germany’s dolus eventualis doctrine sets a lower bar than Thailand’s virtual-certainty requirement, and that Thai law has not adopted that lower threshold. If you’re used to thinking in Japanese-law categories, resist the instinct to map “未必の故意” directly onto Thailand’s indirect intention — they’re not calibrated the same way.

The five-step common-law mens rea ladder — and why Thailand only has two rungs

The professor sketched the mens rea ladder familiar from common-law systems: strict liability → negligence → recklessness → knowledge → purpose, with recklessness sitting in the middle as “you know there’s a risk and you take it anyway, hoping you’ll get away with it” — a step below full knowledge, a step above mere carelessness.

Thai law, he stressed, does not have all five rungs. It has essentially two: intention (direct and indirect, i.e. desire and virtual-certainty-with-indifference) and negligence. There is no independent category of recklessness in Thai criminal law. This matters practically: arguments imported from common-law jurisdictions that hinge on “recklessness” as a distinct, intermediate mental state simply don’t have a home in the Thai statutory scheme — you have to route the argument through negligence instead, even where it feels like it should sit somewhere higher.

The four elements of negligence — and the doctor’s twenty-five seconds

Because intention is so hard to prove in AI cases (as the analysis above shows), negligence under Section 59, paragraph 4 becomes the operative framework in nearly every realistic AI-harm scenario. The professor broke negligence down into four elements, illustrated with a personal, slightly self-deprecating example about his own billing rate as a lawyer to make the “standard of care” concept concrete:

Standard of care — benchmarked against an imaginary “careful person” who shares the same training, experience, age, and position as the actual defendant. A more experienced, more senior professional is held to a higher standard, not because the law treats them as more valuable, but because their greater experience makes a given error less excusable.
Capacity — the defendant’s own internal characteristics: are they a specialist or a general practitioner, newly qualified or a veteran?
Circumstances — external factors at the time: workload, time pressure, how many other patients that day, and (directly on point) a 25-second response window.
Situation — a category the professor acknowledged overlaps somewhat with “circumstances” in practice; he didn’t draw a crisp line between the two live in class, and I’d flag this as worth confirming directly with him before an exam.

He then put the sharpest question of the entire session directly to the class, and — notably — flagged it himself as exam material:

“If the doctor had only twenty-five seconds and no explanation on the screen, what would be meaningful enough to call control? Should we criminally blame the doctor?… There is something I intend to ask you. One of the options I have in the exam.”

Working through this with the analytical tools above: the doctor plainly satisfies elements 1 (there’s a professional standard of care) and 3/4 (the circumstances — 25 seconds, no explanatory information on screen — are established facts). The real fight is over capacity: did the doctor have the realistic capacity to exercise meaningful judgment in 25 seconds, with no explanation of the AI’s reasoning displayed? If capacity is negated, the negligence analysis fails at that element regardless of how the other three come out — recall from Section 5 that capacity is also a necessary element of omission liability, so this single factual point (25 seconds, no explanation) is doing double analytical work, defeating both the omission-capacity element and the negligence-capacity element simultaneously.

This connects to a related framework the professor raised earlier in the course of discussing the EV/pedestrian hypothetical: even where a human is nominally “in the loop,” that doesn’t mean the human has meaningful control. His own illustration made the point vividly — even a driver who is theoretically in control of the car, hands near the wheel, cannot necessarily do anything in the split second an autonomous system swerves to avoid one hazard and creates another. “There’s nothing that can stop the car from hitting that poor woman,” even though a human was, in some formal sense, “in the car.” Keeping a human in the loop is not the same as giving that human the practical capacity to intervene — and criminal negligence liability cannot rest on the mere formal presence of a human where the practical capacity to act is absent.

The standard of care problem specific to AI

One further wrinkle: what standard of care should apply to the AI system itself, when we’re assessing whether deploying it in the first place was negligent? A student proposed that price and stated technological capability should factor in. The professor agreed, but went further, making an important practical concession: lawyers, judges, and regulators cannot realistically verify an AI system’s actual technical capability. No judge is equipped to evaluate a model’s true error rate or robustness. So — pragmatically — the standard of care ends up anchored to what the provider claims in its marketing, documentation, and advertised specifications. If a company advertises that its system performs a task “faster” or “more reliably” or “more logically” than the alternative, the law has little choice but to take that claim at face value when assessing whether reliance on it was reasonable. This has an obvious and slightly uncomfortable implication for anyone deploying AI commercially: your own marketing materials may end up functioning as the evidentiary benchmark against which your negligence is later measured.

8. The Responsibility Gap

The professor introduced Andreas Matthias’s 2004 concept of the “responsibility gap” — the circumstance in which a learning system is developed to operate in ways such that no human retains sufficient control or foreseeability to bear responsibility for its outputs in the conventional sense. He broke this into three practical dimensions:

Material / compensation gap — the injured party cannot obtain compensation through ordinary channels. His view: this is generally addressable through product liability, insurance, or a compensation fund, and does not require inventing a fictional mens rea for the AI system to solve.
Evidential gap — the evidence needed to establish fault simply cannot be obtained (a genuinely hard problem given the “black box” nature of many AI systems).
Capability / control gap — nobody in the causal chain had sufficient practical control to have prevented the outcome.

On this third gap, he was unusually emphatic, warning against the temptation to lower the criminal liability threshold just because AI makes fault-finding difficult:

“If you lower the threshold, criminal law means nothing… Right now, criminal law works as a moral compass for everyone to know what you can, what you cannot do… If you [fracture] the criminal law to accommodate AI, you will lose everything.”

He also referenced (and I want to flag this citation honestly as uncertain — the name came through the recording as something like “Lima,” in a 2018 South Carolina Law Review article, but I was not able to confirm the spelling or the precise title from audio alone, and I’ll update this once I’ve confirmed it directly) an argument that criminal lawyers must be willing to accept that some accidents happen even though no one was negligent — what the article apparently called “bad luck.” His gloss on this, which I think is the single most important normative sentence of the whole session:

“Bad luck may mean no crime. But it does not mean there is no remedy.”

That’s a genuinely useful sentence to carry around outside the classroom too. The absence of criminal liability is not the absence of accountability. Product liability, insurance, regulatory fines, and civil compensation can all do real work even where the criminal law, correctly, declines to reach.

9. Corporate Liability and Its Limits

The gas explosion case, and what it proves

Returning to the corporate-liability example from Section 4: the professor’s point in raising a case where a corporation was fined for negligent homicide following dozens of deaths was not primarily about that case’s outcome (which he plainly regarded as disproportionately lenient — a modest fine against a death toll in the dozens). It was to establish a structural fact: Thailand’s criminal law already has a working mechanism for prosecuting juristic persons. The mechanism (fines, because imprisonment is impossible for an entity with no body) already exists. So the obstacle to holding an AI’s operating company criminally liable for AI-caused harm is not the absence of a corporate-liability mechanism — it’s the prior, separate question of whether the individual elements of the offence (act, duty, causation, fault) can be made out against the corporation on the specific facts.

Uber’s Tempe crash, and the “moral crumple zone”

The professor’s central illustrative case for corporate liability’s limits was the 2018 Uber autonomous-vehicle fatality in Tempe, Arizona — the case in which a self-driving test vehicle struck and killed a pedestrian, Elaine Herzberg.

The facts as he presented them:

The U.S. National Transportation Safety Board (NTSB), roughly a year after the crash, concluded the probable cause was the vehicle operator’s failure to monitor the driving environment, because she was visually distracted by her personal mobile phone during the trip.
But the same report identified this as only one factor among several systemic contributing factors: inadequate safety risk assessment by the technology developer, ineffective oversight of vehicle operators, and a lack of adequate response to known automation-complacency risk (the tendency of humans to disengage when a system usually performs well).
In March 2019, prosecutors concluded that the evidence did not support criminal liability against the Uber corporation itself. The matter was referred back for consideration of charges against the individual vehicle operator only.
The operator ultimately pled to an endangerment offence and received three years of supervised probation. The corporation was never criminally prosecuted, despite being formally identified as a systemic contributor to the crash.

This asymmetry — multiple systemic, organisational causes identified, but only the individual human operator prosecuted — is precisely what the academic literature calls a “moral crumple zone.” Just as a car’s crumple zone is engineered to absorb a crash’s physical force so the passenger compartment doesn’t have to, individual human operators positioned at the point of final interaction with an automated system tend to absorb the legal and moral force of a systemic failure, even when the underlying causes lie further upstream, distributed across the organisation.

The professor’s use of this case was, I think, a deliberate mirror of the hospital hypothetical: the doctor with 25 seconds is Uber’s operator with a distracted moment. Both are the last human touchpoint in a system whose deeper failures were designed, deployed, and maintained by people much further removed from the moment of harm — and in both cases, there’s a real risk that the legal system, for lack of a better mechanism, concentrates blame on the person standing closest to the outcome rather than the organisations that built and ran the underlying conditions.

The missing piece: organisational fault

What both the gas explosion case and Uber’s Tempe crash reveal, when read together, is a structural gap shared by Thai law: there is no general offence of organisational negligence — no crime that directly targets “the corporation’s systems, culture, and processes were negligently designed,” as distinct from “a specific human employee was negligent.” Corporate criminal liability, where it exists, tends to piggyback on an individual’s fault being attributed upward, or on narrow, sector-specific strict/quasi-strict liability provisions. A genuinely organisational failure — a company that adopts an AI system, receives a near-miss warning, and does nothing — can fall into a gap between “no individual employee was clearly negligent enough to convict” and “there’s no free-standing crime of negligent organisational design” to catch the company itself.

(Comparative note: Japanese law has the same structural gap, addressed only partially through 両罰規定 — dual-punishment provisions found in specific administrative-criminal statutes (the Antimonopoly Act, the Food Sanitation Act, and others), which allow a company to be punished alongside a culpable individual employee, but which still require an identifiable individual’s violation as the trigger. Neither Thailand nor Japan has a general, free-standing offence of organisational negligence untethered from an individual’s fault.)

10. A Note on Comparative Law: Where Japan Fits In

I’m Japanese, and one running habit in this LL.M. — encouraged, frankly, by the professor’s own tendency to cold-call students and ask “how is this handled in your country?” — is to keep a mental map of where Thai doctrine and Japanese doctrine line up, and where they don’t. A few of the clearest correspondences from this session:

Thai law concept (this session)	Japanese law equivalent	Note
Legality principle (Penal Code s.2, Constitution s.29)	罪刑法定主義 (Constitution Art. 31, ban on analogical interpretation)	Both systems block extending “person” to cover AI by analogy, regardless of how serious the harm
Acting through an instrument	間接正犯 (indirect perpetration)	Thailand has no single codified general provision for this; Japan has developed it more fully through doctrine, but the underlying idea — using a non-culpable instrument to commit an offence — is the same
Negligence under s.59(4)	業務上過失致死傷罪, Art. 211 (professional negligence causing death/injury)	Japan’s negligence structure (foreseeability, duty of foresight, avoidability, duty of avoidance, negligence, causation) maps closely onto the professor’s four elements (standard of care / capacity / circumstances / situation)
Virtual-certainty indirect intention	確定的故意 more than 未必の故意	Important: don’t equate Thailand’s high “virtual certainty” bar with Japan’s lower-threshold dolus eventualis (未必の故意, which requires only foreseeing a possibility and accepting/認容 it)
No general causation statute; condition theory + s.63 “ordinary consequence”	No general causation statute; 相当因果関係説 / 危険の現実化 (adequate causation / realisation of the danger, both case-law-developed)	Structurally identical two-step architecture (factual causation, then an objective-attribution filter), reached by different doctrinal routes
Missing organisational-fault offence	両罰規定 (dual-punishment provisions in specific administrative-criminal statutes)	Both systems can punish a corporation, but only by piggybacking on an identifiable individual’s violation — neither has a free-standing offence of organisational negligence
Conscious/act vs. conscience/responsibility	行為 (act, requiring voluntary bodily movement) vs. 責任能力 (capacity for criminal responsibility, Art. 39 insanity provisions — 是非弁別能力 and 行動制御能力)	The Thai “conscious/conscience” split and the Japanese act/responsibility-capacity split are doing essentially the same conceptual work

The recurring theme across nearly every row of this table is that Thailand and Japan hit the same structural walls when they try to extend criminal law to AI — no general causation statute, no free-standing organisational-negligence offence, a legality principle that blocks analogical extension of “person” — even though the two systems got there via different legal histories (Thailand’s Penal Code drawing on both civil-law and some common-law influences; Japan’s built on a more purely continental European, primarily German, dogmatic tradition). That convergence, to me, is actually a pretty strong signal that these are genuine structural features of how criminal law works, rather than idiosyncrasies of either jurisdiction — which in turn suggests that whatever eventually gets built to close the AI responsibility gap will probably have to be a fairly general solution, not a Thailand-specific or Japan-specific patch.

11. A Digression: How Words Shape the Analysis

I want to isolate one small point from Section 2 because I think it’s easy to read past it, and I don’t want to.

The very first thing the professor corrected in the entire session was a student saying the hospital “hired” the AI. Not the legal conclusion — the word choice. And the reason that correction mattered, and kept mattering for the rest of the session, is that the vocabulary you use to describe an AI system pre-loads assumptions about its legal status. “Hire” presupposes an employee. “Decide” presupposes a decision-maker. “Know” presupposes a knower. Once those words are in play, a whole apparatus of legal consequences — vicarious liability, mens rea, duty — starts attaching itself almost automatically, before anyone has actually argued for AI personhood on the merits.

This is, I think, a genuinely portable lesson well beyond the hospital hypothetical, and beyond Thai law. If you’re a practitioner drafting an AI vendor contract, a compliance officer writing an internal AI-use policy, or a journalist covering an AI-related harm, the discipline of writing “the company deployed a system that recommended X” instead of “the AI decided X” is not pedantry. It’s the difference between correctly locating legal responsibility in the humans and organisations who built, deployed, and maintained the system, and inadvertently writing your way into a conclusion — “the AI is responsible” — that neither Thai law nor (per Section 10) Japanese law is currently prepared to recognise.

12. Closing Thought

The professor’s framework, distilled to its simplest form, treats “AI and criminal liability” not as one question but as a sequence of ordinary criminal-law questions, asked in order, about a genuinely new and unusually opaque kind of instrument:

Was there an act, or an omission — and if an omission, was there duty and capacity?
What offence, specifically, is even on the table?
Was there a legally sufficient causal link between the conduct and the result?
Did the actor’s state of mind reach intention (with its demanding virtual-certainty threshold for the indirect form), or only negligence — or neither?
If negligence: did the actor have the capacity, in the real circumstances they faced, to meet the applicable standard of care?
And finally, if no individual satisfies all of the above: is there a remedy anywhere else — civil liability, insurance, product liability — even where the criminal law correctly declines to reach?

None of this is really a question of “can AI commit a crime.” It’s a question of how much responsibility the humans who built, sold, deployed, and operated the system actually had — and whether the law currently has good enough tools to trace that responsibility accurately across a chain of many hands. The professor’s own answer, stated at the outset and never really abandoned across two and a half hours of increasingly granular analysis, is that the tools mostly do exist already — they’re just the ordinary tools of duty, act/omission, causation, and mens rea, applied carefully, one actor at a time, resisting the shortcut of either “the AI did it” or “everyone did it, so everyone’s guilty.”

What’s missing isn’t a new category of crime. It’s better evidence, better tracing of causal contribution across the many-hands problem, and — per Uber’s Tempe crash — a serious, unresolved question about why individual operators keep absorbing legal blame that the evidence itself shows was substantially systemic.