Philosophy
"A Computer Can Never Be Held Accountable": The Story Behind the 1979 Slide, and What It Means for AI Agents
The famous slide said to come from 1979 IBM training: what is confirmed about its origin and what is not, the argument behind its two lines, the responsibility gap, what GDPR and the EU AI Act require today, and why it matters more now that AI agents can act on their own.
Akmal Alif · 11 October 2026 MYT
Philosophy is a new EXEPERT section for the questions underneath the technology: who is responsible, what counts as judgment, and what machines should and should not be trusted to do. Builders run into these questions every day, usually without the vocabulary to name them. This section is for slowing down and naming them.
The first post starts with two lines that have become one of the most shared images in AI discussions. In July 2026 the history account @timetphoto posted them to Instagram, describing them as a warning from a 1979 IBM training manual and hedging, rightly, with the word "reportedly" (timetphoto, 2026). The slide reads:
A computer can never be held accountable. Therefore a computer must never make a management decision.
This post covers where the slide comes from, what is and is not known about it, the argument it makes, and why it has become more relevant, not less, now that AI agents can act on their own.
Where the slide comes from
The provenance is thinner than the slide's fame suggests. Simon Willison, who has written about it several times, collected most of the trail (Willison, 2025):
February 2017. The earliest known online appearance is a photo posted on Twitter by the user @bumblebike in reply to another user's thread, describing it as being from a 1979 presentation.
April 2021. The same user, asked about it again, said it came from an internal IBM training session in 1979.
December 2021. The user said the original had been destroyed in a flood in 2019.
December 2022. MIT CSAIL posted the image as an IBM slide from 1979, and it reached millions of people.
November 2023. Jonty Wareing shared a reply from IBM's Corporate Archives, which had searched its collection several times and could not find the presentation. In a later exchange with Willison, Wareing passed on the archive's point that branch offices often produced material that was never archived.
So what is confirmed is that the photo existed by 2017, that the person who posted it attributed it to internal IBM training in 1979, and that IBM's own archive could not locate it. What is not confirmed is that IBM produced it, the year, who wrote it, or which course it belonged to.
The wording has also drifted. The earliest descriptions say "presentation"; many reposts, including the July 2026 one, say "training manual". The photo shows a single hole-punched sheet, which fits a page from a presentation pack better than a bound manual.
None of this makes the slide less useful. Its argument stands on its own. It just means it should be quoted as a reported 1979 IBM slide, not as a documented corporate policy.
The argument
The slide is a two-step syllogism, and both steps are worth taking seriously.
Step one: a computer can never be held accountable. Being accountable means more than being the cause of something. It means being answerable for it: able to explain the decision, to be blamed or sanctioned, to make amends, and to change as a result. A computer can be switched off, patched or replaced, but none of that is accountability. It cannot be fired, sued, disbarred or ashamed. Its maker or operator can be, which is exactly the point: responsibility has to land on a person.
Step two: therefore it must never make a management decision. In 1979 a "management decision" meant a decision about people, money and direction: who to hire or let go, what to fund, what risk to take. Those decisions carry consequences for others, and the people affected are owed someone who answers for them. The slide does not say computers should not help. It says the decision has to belong to someone who can be held to account.
The responsibility gap
Philosophers have given the slide's worry a name. In 2004 Andreas Matthias argued that learning machines create a "responsibility gap". Traditionally the manufacturer or operator of a machine is responsible for what it does. But a system that learns from data can behave in ways its makers could not predict even in principle. Society then faces a choice between giving up such machines and accepting a gap that traditional ideas of responsibility cannot bridge (Matthias, 2004).
Later work refined the idea. Filippo Santoni de Sio and Giulio Mecacci argue that there is not one gap but at least four connected ones: culpability (who deserves blame), moral accountability (who must explain), public accountability (who answers to the public and its institutions) and active responsibility (who should take care in advance). They also show that gaps arise from organisations and law, not only from the technology, and that they can appear even with systems that do not learn. Their proposed answer is "meaningful human control", designed into the whole socio-technical system rather than added as a final check (Santoni de Sio & Mecacci, 2021).
When people trust the output anyway
The slide warns against letting computers decide. A quieter failure is letting a computer's output go unquestioned while a human formally decides.
The Post Office Horizon scandal in the United Kingdom is the starkest recent example. Subpostmasters were prosecuted on the strength of shortfalls reported by the Horizon accounting system, which was later shown to be unreliable. A statutory public inquiry chaired by Sir Wyn Williams published the first volume of its final report in July 2025 (Post Office Horizon IT Inquiry, 2025). In January 2025 the Ministry of Justice opened a call for evidence on the legal presumption, in English law, that computer evidence is reliable unless shown otherwise (Ministry of Justice, 2025). A person signed every charge, but in practice the system's numbers were treated as beyond challenge.
What the law says now
Two pieces of European law turn the slide's principle into rules.
GDPR Article 22 gives people the right not to be subject to a decision based solely on automated processing, including profiling, when it has legal or similarly significant effects on them. There are exceptions, such as when the decision is needed for a contract or the person has given explicit consent, but even then the person keeps at least the right to human intervention, to give their view, and to contest the decision (Regulation (EU) 2016/679, 2016).
The EU AI Act, Article 14 requires high-risk AI systems to be designed so that people can effectively oversee them while they are in use. The people overseeing a system must be able to understand its limits, stay aware of automation bias (the tendency to over-rely on a system's output), interpret its output correctly, disregard or reverse it, and stop the system (Regulation (EU) 2024/1689, 2024). The obligations for high-risk systems have since been postponed: under the 2026 Digital Omnibus amendments they apply from December 2027 or August 2028, depending on the category of system (Lewis Silkin, 2026).
Both rules accept that machines will be involved in decisions. What they insist on is that a person can step in, and that oversight is real, not a rubber stamp.
What it means for AI agents
The slide was written for computers that calculated and reported. Today's AI agents send emails, merge code, book travel and move money. That makes the slide's question practical.
Simon Willison put it in organisational terms in July 2026. Many companies name a "directly responsible individual" for each project, the person ultimately accountable for its success or failure. Willison argues that an agent should never be that person, because humans can take accountability for their actions and machines cannot (Willison, 2026).
For anyone building or deploying agents, that suggests a few habits:
Name an owner. Every agent, and every workflow it runs, should have a named person who answers for what it does.
Match autonomy to consequence. Let agents act freely where mistakes are cheap and reversible. Require human approval before anything that is costly, public or hard to undo.
Make the reasoning inspectable. Keep a record of what the agent saw, what it decided and why, so the owner can actually explain a decision.
Make overriding easy. A stop button and an undo path matter more than a confidence score.
Watch for automation bias. Approving every suggestion without reading it is not oversight. If reviewers never disagree, the review is not working.
The slide's two lines are short enough to fit on a sticky note, and they still describe the right default. Computers can do more of the work every year. Somebody still has to own the decision.
References
Lewis Silkin. (2026, July 27). The Digital Omnibus on AI enters into force today. https://www.lewissilkin.com/en/insights/2026/07/27/the-digital-omnibus-on-ai-enters-into-force-today-102nedo
Matthias, A. (2004). The responsibility gap: Ascribing responsibility for the actions of learning automata. Ethics and Information Technology, 6(3), 175–183. https://doi.org/10.1007/s10676-004-3422-1
Ministry of Justice. (2025, January 21). Use of evidence generated by software in criminal proceedings [Call for evidence]. GOV.UK. https://www.gov.uk/government/calls-for-evidence/use-of-evidence-generated-by-software-in-criminal-proceedings
Post Office Horizon IT Inquiry. (2025, July 8). Volume 1 of the Post Office Horizon IT Inquiry's final report. https://www.postofficehorizoninquiry.org.uk/volume-1-post-office-horizon-it-inquirys-final-report
Regulation (EU) 2016/679. (2016). General Data Protection Regulation, Article 22: Automated individual decision-making, including profiling. Official Journal of the European Union, L 119. https://eur-lex.europa.eu/eli/reg/2016/679/oj
Regulation (EU) 2024/1689. (2024). Artificial Intelligence Act, Article 14: Human oversight. Official Journal of the European Union, L series. https://eur-lex.europa.eu/eli/reg/2024/1689/oj
Santoni de Sio, F., & Mecacci, G. (2021). Four responsibility gaps with artificial intelligence: Why they matter and how to address them. Philosophy & Technology, 34(4), 1057–1084. https://doi.org/10.1007/s13347-021-00450-x
timetphoto. (2026, July 18). An IBM training manual from 1979 reportedly warned… [Instagram post]. Instagram. https://www.instagram.com/p/Dba5BPBAq2v/
Willison, S. (2025, February 3). A computer can never be held accountable. Simon Willison's Weblog. https://simonwillison.net/2025/Feb/3/a-computer-can-never-be-held-accountable/
Willison, S. (2026, July 12). Directly responsible individuals (DRI). Simon Willison's Weblog. https://simonwillison.net/2026/Jul/12/directly-responsible-individuals/