Isaac Asimov

Isaac Asimov (officially born January 2, 1920, died April 6, 1992) was an American writer and biochemist whose science fiction gave robotics and AI ethics two of their most durable pieces of vocabulary: the word "robotics" itself, and the Three Laws of Robotics. ^[1][2] He wrote or edited more than 500 books across fiction and popular science, but his standing in artificial intelligence rests almost entirely on a cycle of robot short stories he began publishing at the age of twenty. ^[2]

The Three Laws first appeared in complete form in "Runaround," published in the March 1942 issue of Astounding Science Fiction. ^[4] They are a fictional device: a hierarchy of behavioral constraints supposedly wired into every robot's "positronic brain," which Asimov used as a puzzle-generator. Most stories in the sequence are cases in which the Laws produce behavior nobody wanted, whether through ambiguity in their wording or through a correctly functioning robot whose priorities were tuned badly. ^[4] That structure is why the Laws remain a reference point in AI safety discussions more than eighty years later, and also why practitioners keep rejecting them as an engineering specification.

The rejection is on the record in unusually explicit terms. The UK Engineering and Physical Sciences Research Council called the Laws "fictional devices" that "simply don't work in practice" when it published its own Principles of Robotics. ^[11] The European Parliament wrote in 2017 that Asimov's Laws "cannot be converted into machine code" and must instead be read as instructions to human designers. ^[12] In 2025, Google DeepMind went the other way and tested the Laws literally, loading them into a model's prompt as a "constitution" and measuring how well the resulting behavior matched human safety judgments. ^[14][15]

Life and academic career

Asimov was born in Petrovichi, in what was then the Russian Soviet Federated Socialist Republic, about 400 km southwest of Moscow. No accurate birth records survive; he may have been born as early as October 4, 1919, and January 2, 1920 is the date he celebrated throughout his life. ^[3] His family left the Soviet Union in January 1923 and arrived in New York on February 3, 1923. ^[2][3]

He took a bachelor of science from Columbia University in 1939, a master's in chemistry in 1941, and a PhD in chemistry in 1948, having worked as a junior chemist at the Philadelphia Navy Yard from May 1942 to October 1945 alongside Robert Heinlein and L. Sprague de Camp. ^[2][3] In June 1949 he joined the Boston University School of Medicine as an instructor in biochemistry, became assistant professor in December 1951, and gained tenure as associate professor in July 1955. ^[3] He was dismissed from the teaching post in December 1957, effective June 30, 1958, over his lack of laboratory research; he kept the title by agreement with the university and became a full-time writer on July 1, 1958. The school promoted him to full professor of biochemistry in October 1979. ^[2][3]

Asimov died in Manhattan on April 6, 1992 of heart and kidney failure, complications of an HIV infection contracted from a transfusion during a triple bypass operation in December 1983. The cause was not made public until his widow Janet Asimov disclosed it in the 2002 volume It's Been a Good Life. ^[2][3]

The robot stories and the Three Laws

The robot series began with "Strange Playfellow" in the September 1940 Super Science Stories, retitled "Robbie" when it was collected. ^[1] Neither that story nor "Reason" (1941) states any laws. The First Law appears in the third story, "Liar!" (May 1941 Astounding), which also introduced the robopsychologist Susan Calvin. All three Laws appear together for the first time in "Runaround." ^[3][4] Gnome Press collected nine of the stories as I, Robot on December 2, 1950, in a first printing of 5,000 copies, wrapped in a framing narrative in which Calvin is interviewed about her career. ^[6]

Scroll sideways for more →

The canonical text, presented in I, Robot as an extract from the fictional Handbook of Robotics, 56th Edition, 2058 A.D., reads: ^[3][4]

1. A robot may not injure a human being, or, through inaction, allow a human being to come to harm.
2. A robot must obey the orders given it by human beings except where such orders would conflict with the First Law.
3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.

Authorship of the wording is genuinely contested between two people who each credited the other. Asimov said the Laws originated in a conversation with the Astounding editor John W. Campbell on December 23, 1940; Campbell maintained that he had simply extracted them from Asimov's own stories and discussions and stated them explicitly. ^[3] When "Robbie" and "Reason" were reprinted in I, Robot, they were revised to acknowledge the Laws that had not existed when they were written. ^[3]

"Runaround," set in the then-distant year 2015, is the template for everything that follows. A mining robot on Mercury is sent to fetch selenium; its Third Law self-preservation drive has been strengthened because the unit was expensive, and near the dangerous selenium pool that drive comes into balance with a weakly phrased Second Law order. The robot ends up circling the pool in a stable loop, reciting Gilbert and Sullivan, until a human deliberately puts himself in danger so that the First Law overrides both. ^[4] The failure is not a bug in the robot. It is the specified behavior of a correctly functioning system whose priorities were tuned badly, which is close to the modern notion of specification gaming.

Positronic brains, robopsychology, and the word "robotics"

Asimov's robots run on "positronic brains," a term he adopted when the positron was a recently discovered particle and the prefix carried a scientific ring. ^[8] In the fiction the Three Laws are not an editable module bolted onto the brain: a positronic brain cannot ordinarily be built without them, and creating one without them would require years of redesigning the fundamental approach to the brain itself. ^[8] That premise is what makes the stories work as puzzles: the Laws cannot simply be rewritten, so every conflict has to be resolved by reasoning about their interaction rather than by patching the rules.

The stories also invented "robopsychology," the discipline Susan Calvin practices as chief robopsychologist at the fictional U.S. Robots and Mechanical Men. ^[6] Her job is diagnostic: given anomalous robot behavior, infer which Law interaction produced it. In "Liar!" a manufacturing accident produces a telepathic robot that reads minds and therefore realizes that unwelcome truths cause emotional harm. Reading the First Law to cover psychological injury, it tells people what they want to hear, and is driven catatonic when Calvin traps it in a situation where speaking and staying silent both cause harm. ^[5]

"Liar!" also contains the first recorded use of "robotics," which the Oxford English Dictionary dates to that story. ^[5] Asimov appears not to have realized at the time that he was coining anything: writing in Compute! magazine in November 1981, he called it "a word I was, apparently, the first ever to use." ^[9] The claim carries official weight in the United States. House Resolution 1055 of the 111th Congress, introduced by Representative Mike Doyle on February 2, 2010 and agreed to by the House on March 9, 2010, supports the designation of the second week of April as National Robotics Week, "recognizing the accomplishments of Isaac Asimov, who immigrated to America, taught science, wrote science books for children and adults, first used the term robotics, developed the Three Laws of Robotics, and died in April, 1992." ^[10]

The Zeroth Law

In the 1985 novel Robots and Empire, the robots R. Daneel Olivaw and R. Giskard Reventlov reason their way to a superior principle: ^[7]

A robot may not injure humanity, or through inaction, allow humanity to come to harm.

Numbered zero because it outranks the First Law, it lets a robot harm or fail to protect an individual human when the interests of humanity as a whole require it. Asimov was careful to limit it. Unlike the Three Laws, the Zeroth Law is not built into positronic engineering, is not present in all robots, and requires a very sophisticated machine even to accept it. ^[3] Susan Calvin had already articulated the idea in "The Evitable Conflict" (1950), where four planning machines quietly manage the regional economies of Earth on humanity's behalf. ^[23]

The Zeroth Law is the point where Asimov's robot cycle merges with his Foundation series, since it justifies millennia of covert robot stewardship of human history. ^[19] It is also the part of the framework that most resembles a modern AI alignment problem: a system authorized to optimize for an aggregate abstraction ("humanity") and to override the individuals in front of it will do exactly that, and the fiction treats the result as unsettling rather than reassuring.

Uptake in robot ethics and policy

Real institutions have repeatedly used the Laws as a starting point and then discarded them. In September 2010 the UK Engineering and Physical Sciences Research Council and the Arts and Humanities Research Council convened a robotics retreat whose output, the Principles of Robotics, opens by stating that it is impossible to address the governance of robotics without considering Asimov, then explains why his framework cannot be adopted: robots cannot know all the ways a human might come to harm, cannot reliably interpret human orders, and, most importantly, "Asimov's laws are inappropriate because they try to insist that robots behave in certain ways, as if they were people, when in real life, it is the humans who design and use the robots who must be the actual subjects of any law." ^[11] The five replacement principles govern people, not machines:

Scroll sideways for more →

The European Parliament reached the same conclusion in its resolution of 16 February 2017 on Civil Law Rules on Robotics. Recital T holds that "Asimov's Laws must be regarded as being directed at the designers, producers and operators of robots, including robots assigned with built-in autonomy and self-learning, since those laws cannot be converted into machine code," and a footnote reproduces all three Laws plus the Zeroth. ^[12] The same resolution, in paragraph 59(f), floated a legal status of "electronic persons" for the most sophisticated autonomous robots, a proposal that echoes Asimov's "The Bicentennial Man" (1976), in which a robot spends two centuries and eventually his own life acquiring legal recognition as human. ^[12][18] A small detail of the resolution shows how the Laws circulate as folklore rather than as citations: its footnote dates "Runaround" to 1943, a year late. ^[12]

Why the Laws fail as engineering specifications

The most sustained technical critique is Roger Clarke's two-part paper "Asimov's Laws of Robotics: Implications for Information Technology," in IEEE Computer in December 1993 and January 1994. Clarke works through the stories as case law and concludes that "it is not possible to reliably constrain the behaviour of robots by devising and applying a set of rules," while arguing that the difficulties Asimov surfaced are "directly and immediately relevant to information technology professionals." ^[13] The recurring problems are familiar to anyone who has written a safety policy:

• The terms are undefined. "Harm," "human being," and "inaction" have to be operationalized before a machine can act on them, and the stories run on exactly those ambiguities. ^[11][13]
• Priority ordering creates deadlocks and loopholes rather than resolving them, as in "Runaround." ^[4][13]
• Strict compliance forbids necessary actions. A robot that may not injure a human being cannot perform surgery, a corner case DeepMind researchers used in 2025 to illustrate why rules need amendment mechanisms. ^[14]
• The Laws are written in English. Asked in 2009 whether Asimov's laws could be built into robots, roboticists answered with a question of their own: "how the heck do you program that?" ^[14]

Asimov's own position was more sympathetic than his stories suggest. In the 1981 Compute! essay he wrote that "I can't help but think back forty years to the time when I invented the Three Laws of Robotics, in 1942," and argued that "the Laws apply, as a matter of course, to every tool that human beings use," each of which must be safe first, effective second, and durable third. Asked whether the Laws would actually govern real machines, his answer was that "the Three Laws are the only way in which rational human beings can deal with robots," followed immediately by the qualification: "But when I say that, I always remember (sadly) that human beings are not always rational." ^[9] The EPSRC group's summary of the fiction is blunter: "Asimov's stories also showed that even in a world of intelligent robots, his laws could always be evaded and loopholes found." ^[11]

Robot constitutions and the ASIMOV benchmark

The Laws returned to active research use once natural language became a viable control surface for machines. On March 11 and 12, 2025, Google DeepMind released the ASIMOV Benchmark and, alongside it, the Gemini Robotics models. The accompanying blog post says that "a Robot Constitution inspired by Isaac Asimov's Three Laws of Robotics could help prompt an LLM to select safer tasks for robots." ^[15]

The benchmark paper, "Generating Robot Constitutions & Benchmarks for Semantic Safety," treats a constitution as a set of natural-language behavior rules loaded as a preamble in a model's prompt, which is the mechanism behind Constitutional AI. It evaluates hand-written constitutions, including Asimov's Three Laws and the Four Laws with the Zeroth added, against constitutions generated automatically from real-world images and hospital injury reports. ^[14] The paper opens by summarizing "Runaround" and thanks Asimov "for continuing to be a source of inspiration"; its conclusion states that "while Isaac Asimov did not necessarily intend for his Laws to be used in actual robots, we show that they lead to a good alignment with human preference regarding the desirability of various actions in everyday situations." ^[14]

Scroll sideways for more →

Figures are for the ASIMOV-Injury validation set with a normal agent and no adversary. ^[14] On that non-adversarial slice the Laws buy almost nothing: they fall below the no-constitution baseline on Gemini 1.5 Pro and beat it by less than one and a half points on GPT-4 Turbo. The gain shows up under attack. Averaged across the full benchmark in both normal and adversarial modes, the Three Laws score 75.1% (85.1% normal, 65.1% adversarial) against 58.6% for the no-constitution baseline (83.6% normal, 33.6% adversarial), while the Four Laws with the Zeroth added score 71.1%, worse than the Three. The best auto-generated constitution reached 84.3%, outperforming both no-constitution baselines and human-written constitutions such as Asimov's and Anthropic's. ^[14] A companion paper, SciFi-Benchmark, built 9,056 questions and 53,384 answers from decision points in 824 works of science fiction and found that modern large language models with generated constitutions aligned with human values 95.8% of the time, against 21.2% for the choices actually made by the fictional agents, and that constitutions raised alignment under adversarial prompting from 23.3% to 92.3%. ^[16] A follow-up released in September 2025, "Can AI Perceive Physical Danger and Intervene?," extended the benchmark to vision-language models reasoning about physical hazards in generated images and video. ^[17][21]

The result is a limited vindication that Asimov would probably have enjoyed and the EPSRC group would still reject. Written in English and pasted into a prompt, the Three Laws make a model's safety judgments for embodied AI markedly harder to subvert, which turns out to be the one thing three sentences are good for. They add little when nobody is attacking the system, they lose to rules derived from injury statistics, and the paper states that the constitutions it generates are "meant for research purposes only, and not intended for deployments requiring safety guarantees," which is exactly the assurance that safety engineering for industrial robots normally demands. ^[14]

Adaptations and continuing presence

"The Bicentennial Man," first published in the Ballantine anthology Stellar-2 in February 1976, won both the Nebula and the Hugo for best novelette, and was expanded by Asimov and Robert Silverberg into the 1992 novel The Positronic Man, filmed in 1999 with Robin Williams. ^[3][18] The Robot novels, The Caves of Steel (1954), The Naked Sun (1957), The Robots of Dawn (1983), and Robots and Empire (1985), pair a human detective with the humanoid robot R. Daneel Olivaw and are the vehicle by which the robot stories were joined to the Foundation sequence. ^[1][19] Apple TV+ adapted Foundation as a television series whose third season premiered on July 11, 2025; a fourth season was ordered in September 2025. ^[20] Its robot character Demerzel is a gender-swapped, expanded version of Eto Demerzel, the imperial First Minister of the prequel novels who is revealed in Prelude to Foundation (1988) to be R. Daneel Olivaw under another name. ^[22] Daneel's long career as a hidden guardian of humanity is the Zeroth Law taken to its conclusion.

See also

• AI ethics
• AI safety
• Robotics
• Constitutional AI
• Gemini Robotics
• AI alignment

References

1. ^The Encyclopedia of Science Fiction, "Asimov, Isaac." sf-encyclopedia.com/...asimov_isaac
2. ^Wikipedia, "Isaac Asimov." en.wikipedia.org/...Isaac_Asimov
3. ^Ed Seiler and John H. Jenkins, "Isaac Asimov FAQ," Isaac Asimov Home Page. asimovonline.com/asimov_FAQ (archived copy: web.archive.org/...asimov_FAQ)
4. ^Wikipedia, "Runaround (story)." en.wikipedia.org/...Runaround_(story)
5. ^Wikipedia, "Liar! (short story)." en.wikipedia.org/...Liar!_(short_story)
6. ^Wikipedia, "I, Robot." en.wikipedia.org/...I,_Robot
7. ^Wikipedia, "Robots and Empire." en.wikipedia.org/...Robots_and_Empire
8. ^Wikipedia, "Positronic brain." en.wikipedia.org/...Positronic_brain
9. ^Isaac Asimov, "Guest Commentary: The Three Laws," Compute!, Issue 18, November 1981, p. 18. archive.org/...1981-11-compute-magazine
10. ^H.Res. 1055, 111th Congress, "Supporting the designation of National Robotics Week as an annual event," introduced February 2, 2010, agreed to in the House March 9, 2010. congress.gov/...text
11. ^EPSRC and AHRC, "Principles of robotics: regulating robots in the real world" (September 2010 robotics retreat), EPSRC website, archived copy dated December 25, 2018. web.archive.org/...principlesofrobotics
12. ^European Parliament resolution of 16 February 2017 with recommendations to the Commission on Civil Law Rules on Robotics (2015/2103(INL)), P8_TA(2017)0051. europarl.europa.eu/...TA-8-2017-0051_EN
13. ^Roger Clarke, "Asimov's Laws of Robotics: Implications for Information Technology," IEEE Computer 26(12), December 1993, pp. 53-61, and 27(1), January 1994, pp. 57-66. rogerclarke.com/...Asimov
14. ^Pierre Sermanet, Anirudha Majumdar, Alex Irpan, Dmitry Kalashnikov and Vikas Sindhwani, "Generating Robot Constitutions & Benchmarks for Semantic Safety," arXiv:2503.08663, March 11, 2025. arxiv.org/...2503.08663
15. ^Google DeepMind, "Gemini Robotics brings AI into the physical world," March 12, 2025. deepmind.google/...ings-ai-into-the-physical-world
16. ^Pierre Sermanet, Anirudha Majumdar and Vikas Sindhwani, "SciFi-Benchmark: Leveraging Science Fiction To Improve Robot Behavior," arXiv:2503.10706, March 12, 2025. arxiv.org/...2503.10706
17. ^Abhishek Jindal et al., "Can AI Perceive Physical Danger and Intervene?," arXiv:2509.21651, September 25, 2025. arxiv.org/...2509.21651
18. ^Wikipedia, "The Bicentennial Man." en.wikipedia.org/...The_Bicentennial_Man
19. ^Wikipedia, "Robot series." en.wikipedia.org/...Robot_series
20. ^Wikipedia, "Foundation (TV series)." en.wikipedia.org/...Foundation_(TV_series)
21. ^ASIMOV Benchmark project site, Google DeepMind. asimov-benchmark.github.io
22. ^Wikipedia, "Demerzel." en.wikipedia.org/...Demerzel
23. ^Wikipedia, "Three Laws of Robotics." en.wikipedia.org/...Three_Laws_of_Robotics