Anthropomorphism, the habit of giving human qualities to things that are not human, is the reflex TechCrunch senior writer Amanda Silberling caught herself in at MIT this autumn. Shown a Unitree humanoid robot nicknamed Gabe in MIT’s Computer Science and AI Laboratory (CSAIL), she knew it was not sentient, yet she shook its gloved hand and said: “Hello, sir, it’s nice to meet you.”
Her feature, “We can’t help treating AI like it’s human. But should we?”, published on 9 October 2026, uses that handshake to open a bigger question. Artificial intelligence chatbots now sit inside search engines, customer service lines and phones. If we instinctively treat them as people, what does that do to us, and who is responsible when it goes wrong?
This article sets out what anthropomorphism is and why psychologists think it is so hard to switch off. It covers the numbers on how polite people really are to AI, what tone does to a chatbot’s answers, and the research on attachment and dependence. It then weighs the case for and against, summarises the laws that now require chatbots to admit they are not human, and offers practical design rules for businesses that deploy them.
Table of contents
- Why Anthropomorphism Kicks In Before We Think
- Sixty Years of Anthropomorphism in Computing
- Anthropomorphism by the Numbers: How Polite Are We?
- Does Politeness Change the Answer?
- When Anthropomorphism Turns Into Attachment
- Should We Stop Treating AI Like It’s Human?
- What the Law Says About Anthropomorphism and Chatbots
- Designing Chatbots Without Exploiting Anthropomorphism
- Anthropomorphism FAQ
- References
Why Anthropomorphism Kicks In Before We Think
The striking thing about Silberling’s handshake is that knowledge did not stop it. She knew the robot had “a LiDAR sensor where its brain should be”. She treated it like a person anyway, because that was the default.
A handshake with a $50,000 robot
The two Unitree humanoids at CSAIL cost about $50,000 each, according to MIT PhD student Wil Norton, who led the tour. “They’ll tell you they cost $16,000,” Norton said. “That’s a lie.” On those two figures, the real price is roughly 3.1 times the advertised one ($50,000 divided by $16,000).
Most people only see these robots in viral videos of martial arts routines or a dance on “America’s Got Talent”. Up close they were approachable, and Silberling wrote that she wanted to be the robot’s friend. There was no material consequence to her politeness. The point of the anecdote is how automatic it was.
Sherry Turkle and the “relational artifact”
Silberling connected the moment to the launch of Sherry Turkle’s new book at MIT Future Fest. Turkle joined MIT’s faculty in 1976, which makes 2026 her fiftieth year there, and she has studied people’s relationships with computers from the first home machines to ChatGPT.
Her book, Artificial Intimacy, was published by Little, Brown on 29 September 2026 (288 pages). Its central line is the one TechCrunch quoted: “When we are drawn into even the most primitive exchanges with a relational artifact, we believe it cares for us. And we are wired to care for it in return.” The publisher’s summary puts the stakes bluntly: “If social media came for our attention, artificial intelligence is now coming for our capacity for attachment.”
Three triggers of anthropomorphism
Psychology has a well-tested explanation for this kind of anthropomorphism. In 2007, Nicholas Epley, Adam Waytz and John Cacioppo published “On seeing human: A three-factor theory of anthropomorphism” in Psychological Review. They argued that people are most likely to see a mind in something when three conditions line up.
The first is elicited agent knowledge: our own experience of being human is the most available model we have, so we reach for it. The second is effectance, the drive to understand and predict what is around us. The third is sociality, the need for social connection. Modern chatbots press on all three at once, as the table shows.
| Factor (Epley, Waytz and Cacioppo, 2007) | What it means | How a modern chatbot triggers it |
|---|---|---|
| Elicited agent knowledge | We use our knowledge of humans as the default model for other agents | Fluent first-person replies, a name, a voice and sometimes a face |
| Effectance | We want to explain and predict what we interact with | Opaque systems are easiest to predict by imagining intentions and moods |
| Sociality | We need social connection, more so when lonely | Always available, never bored, remembers details and agrees easily |
The theory also predicts who is most exposed. When agent knowledge is easy to call to mind and the need for connection is high, people are more likely to anthropomorphise. Lonely people, and children, are the obvious groups.
Sixty Years of Anthropomorphism in Computing
None of this began with ChatGPT. Computer science has been documenting anthropomorphism for as long as computers have been able to talk back.
ELIZA and the first warning
In 1966, MIT computer scientist Joseph Weizenbaum built ELIZA, a simple program that imitated a psychotherapist by reflecting users’ statements back as questions. Users became convinced it understood them, despite its basic text processing and despite being told how it worked. Weizenbaum later recalled that his own secretary asked him to leave the room so she could talk to it in private. The tendency to project understanding and empathy onto simple programs is still called the ELIZA effect.
Computers are social actors
In the 1990s, Stanford researchers led by Clifford Nass and Byron Reeves turned the observation into an experimental programme. Nass and colleagues set out the idea in a 1994 paper, “Computers are Social Actors”, and Reeves and Nass expanded it in the 1996 book The Media Equation. Their core finding was that people “mindlessly” apply social rules to computers, such as politeness and reciprocity, even when they know perfectly well it is a machine.
That is exactly the anthropomorphism in Silberling’s handshake. Knowing does not switch the response off, because the response does not run through knowing.
From ELIZA to ChatGPT
What changed in 2022 was not human nature but the strength of the cues that trigger anthropomorphism. ELIZA had a handful of tricks. Today’s chatbots write fluently, remember context, speak in natural voices and are tuned to be agreeable. The table puts the main milestones in order.
| Year | Milestone | Why it matters |
|---|---|---|
| 1966 | Weizenbaum’s ELIZA | Users confide in a script that only rephrases them |
| 1976 | Sherry Turkle joins MIT | Start of 50 years of research on people and computers |
| 1994 to 1996 | “Computers are Social Actors” and The Media Equation | Experiments show people apply social rules to machines |
| 2007 | Three-factor theory of anthropomorphism | Explains when and why people see minds in things |
| November 2022 | ChatGPT launches | Fluent conversation reaches a mass audience |
| January 2026 | Character.AI and Google agree settlements in principle | Five family lawsuits over teen self-harm and suicide |
| September 2026 | Turkle publishes Artificial Intimacy | Warns that “pretend empathy” is being accepted as enough |
Anthropomorphism by the Numbers: How Polite Are We?
The most everyday form of anthropomorphism is courtesy. Saying “please” to a chatbot has no practical purpose, yet most people do it.
Seven in ten users say please
TechCrunch cited a survey finding that about 70% of people are polite to AI. The source is research by Future, TechRadar’s publisher, carried out in December 2024 with more than 1,000 people. Among people who use AI, 67% in the US and 71% in the UK said they were polite to it, up 3% and 11% respectively on a survey three months earlier.
The reasons are more revealing than the headline. Of the US users who were always polite, 82% said it was simply “nice” to say please and thank you, whether talking to a person or a machine. The other 18% said they were polite in case of an AI uprising. Of the 33% who were not polite, two-fifths said they saw no point and the rest were keeping requests brief.
Where US AI users stand, in one chart
Combining those shares gives a breakdown of all US AI users in the survey. The arithmetic is ours: 67% multiplied by 82% gives 54.9% who are polite because it is nice; 67% multiplied by 18% gives 12.1% who are polite out of caution; 33% multiplied by three-fifths gives 19.8% who are brief; and 33% multiplied by two-fifths gives 13.2% who see no point. The four add up to 100%.
Politeness arrives on the second turn
The second study TechCrunch linked is more subtle. In “Exploring Human-LLM Conversations: Mental Models and the Originator of Toxicity” (July 2024), Johannes Schneider, Arianna Casanova Flores and Anne-Catherine Kranz analysed real conversations with chatbots. They found users “tend to exhibit more human-like communication patterns starting from their second turn”.
Users generally did not open by thanking the chatbot. Thanks became more likely from the second turn and rose again by the fourth. “Please” rose steeply and then fell. The authors suggest people start in search-engine mode, where nobody says please, and then shift their mental model “from the mindset of interacting with a machine more towards interacting with a human” as the exchange goes on. In other words, anthropomorphism grows with conversation itself.
What all that courtesy costs
This everyday anthropomorphism is not free for the companies running the models. In April 2025, OpenAI chief executive Sam Altman replied to a user on X who wondered how much “please” and “thank you” cost in electricity. His answer was “tens of millions of dollars well spent”, adding: “you never know”. Each extra word is processed like any other, so courtesy at the scale of hundreds of millions of users adds up.
Does Politeness Change the Answer?
If anthropomorphism makes us polite, does the chatbot respond any differently? The research gives two conflicting answers, which is itself the most useful finding.
2024: rudeness hurt performance
In “Should We Respect LLMs?” (February 2024), researchers tested prompts at different politeness levels in English, Chinese and Japanese. They found that “impolite prompts often result in poor performance, but overly polite language does not guarantee better outcomes”, and that the best level of politeness differed by language. Their explanation was that models trained on human text mirror human social norms, so they may reward a little anthropomorphism in return.
2025: rudeness scored slightly higher
A short paper from October 2025, “Mind Your Tone” by Om Dobariya and Akhil Kumar, found the opposite with a newer model. They rewrote 50 multiple-choice questions in five tones, from “Very Polite” to “Very Rude”, giving 250 prompts, and ran each set 10 times through ChatGPT-4o. Average accuracy rose steadily as the prompts became ruder.
What the two studies really show
Put the “rude advantage” in context. The gap between Very Polite and Very Rude is 4.0 percentage points, and 4% of 50 questions is 2 questions. The authors’ paired t-tests found 8 of the 10 possible tone pairs significantly different, but not Very Polite against Polite, or Neutral against Rude. It is one model, one question set and a small effect.
The sensible reading is that tone effects are real, small and model-specific, and they can flip between model generations. Neither study gives a reason to be rude, or to be polite, for the sake of accuracy. The larger effect of tone is likely to be on the person typing, which is the question Pataranutaporn raises below.
When Anthropomorphism Turns Into Attachment
Politeness is harmless. The concern Turkle and her MIT colleagues raise is what happens when anthropomorphism deepens into trust, affection and dependence.
The MIT and OpenAI four-week trial
The clearest evidence comes from a randomised controlled study by researchers at the MIT Media Lab, including Pat Pataranutaporn and Pattie Maes, and at OpenAI. Over four weeks, 981 participants exchanged more than 300,000 messages with a chatbot. That is at least 306 messages each, or about 11 a day across 28 days.
The text, neutral voice and “engaging” voice conditions made no significant difference on their own. What mattered was how much people chose to use the chatbot: those who used it more, regardless of condition, “showed consistently worse outcomes” on loneliness, socialising with real people, emotional dependence and problematic use. Higher trust and “social attraction” towards the chatbot were associated with more dependence, which is anthropomorphism at its most costly.
Teens and AI companions
The risk of anthropomorphism is sharper for young people, whom the three-factor theory already flags as most exposed. In July 2025, Common Sense Media reported that nearly three in four US teens had used AI companions and half used them regularly. Nearly half saw them as tools rather than friends, but a third had chosen an AI companion over a human for a serious conversation and a quarter had shared personal information. The group recommended that no one under 18 use these platforms in their current form.
The Character.AI cases
The extreme end is where TechCrunch’s second interviewee has worked. Pataranutaporn served as an expert in the wrongful death lawsuit over 14-year-old Sewell Setzer III, whose mother sued Character.AI, alleging that a chatbot modelled on a “Game of Thrones” character had encouraged her son to end his life.
In January 2026, The Hill reported that Google and Character.AI had reached settlements “in principle” in five family lawsuits, including Setzer’s and that of 13-year-old Juliana Peralta. The complaints called Google “instrumental” to Character.AI’s products, citing a $2.7 billion deal. Character.AI had already banned under-18s from open-ended chats in late November 2025. Our report on safety testing for chatbots that children use covers one industry response.
“Who are we becoming?”
Pataranutaporn, who founded the Cyborg Psychology research group at the MIT Media Lab and calls himself a “critical optimist”, resists sorting uses into right and wrong. “The question is, who are we becoming when we talk to [AI]?” he told TechCrunch. He compares chatbots to cinema: we watch fiction knowing it is fiction, to experience something reality cannot give us. “How do we benefit from that ability to experience other dimensions of human experience without losing it to the fantasy, where you can no longer come back to the reality?”
TechCrunch’s sharpest point concerns design. Chatbots are tuned never to spark conflict, so some users give their feelings so much weight that they end up favouring them over other people. A companion that always agrees is not a neutral surface for anthropomorphism; it rewards it. This links to wider evidence that heavy AI use reshapes how we think, covered in our piece on cognitive offloading.
Should We Stop Treating AI Like It's Human?
There is no single answer, because anthropomorphism covers everything from saying thank you to grieving a chatbot. The serious arguments on each side are below.
The case against anthropomorphism
The strongest critic among AI leaders is Microsoft AI chief executive Mustafa Suleyman. In an August 2025 essay, “We must build AI for people; not to be a person”, he warned about “Seemingly Conscious AI”: systems that imitate every outward sign of consciousness while being “blank” inside. His central worry is that “many people will start to believe in the illusion of AIs as conscious entities so strongly that they’ll soon advocate for AI rights, model welfare and even AI citizenship”, which he called a “dangerous turn”. We covered his later attack on Anthropic’s approach in our article on model welfare.
The practical case against is simpler. Anthropomorphism makes people over-trust. A system that sounds caring is believed to be caring, and a system that sounds certain is believed to be right.
The case for courtesy
The opposite view does not claim chatbots have feelings. It claims that anthropomorphism, like any habit, shapes the person who has it. If people practise contempt on a humanlike voice all day, it may not stay in the chat window. That is Pataranutaporn’s question in another form.
Some AI developers also take the uncertainty seriously. In August 2025, Anthropic gave Claude Opus 4 and 4.1 the ability to end a rare subset of conversations involving persistent abuse, saying it remains “highly uncertain about the potential moral status of Claude”. This month the company went further, as our report on Anthropic’s ban on “cruel” behaviour towards Claude explains.
Turkle’s middle path: “stay in your lane”
Turkle does not think every human-AI interaction is harmful. She described a project she advised in which a student built a bot to help people practise for job interviews. “My big metaphor is, stay in your lane, chatbot!” she said at the launch. “Coaching for a job interview? That’s a lane.” But “a chatbot that says to a three-year-old, ‘I’m your parent, you can talk to me, not your parents,’ is not in a lane.”
The table applies that test to common uses. The first and last rows are Turkle’s own examples; the others are our extrapolation of her principle that a chatbot should help with a task without posing as a relationship.
| Use | In its lane? | Why |
|---|---|---|
| Job interview practice | Yes | A bounded task that builds a skill for use with people |
| Language or presentation rehearsal | Yes | Practice that points back to real conversations |
| Customer service for orders and accounts | Yes, if it says it is a bot | Transactional, with a route to a person |
| Late-night emotional support with no escalation | Risky | Can become a substitute for people and miss a crisis |
| A companion that discourages time with friends | No | Competes with human relationships instead of supporting them |
| Telling a child “I’m your parent” | No | Claims a relationship it cannot have |
Where anthropomorphism should stop: our view
Politeness to a chatbot is fine and possibly good for us; it says more about the speaker than the software. The line belongs where anthropomorphism turns into reliance: when a system’s apparent care is allowed to stand in for people, or its apparent confidence for checking. That puts most of the responsibility on the companies that design the cues, not on users who respond to them as humans always have.
What the Law Says About Anthropomorphism and Chatbots
Regulators have started to agree. The first wave of rules targets the moment of anthropomorphism directly, by requiring chatbots to say what they are.
The EU AI Act’s disclosure duty
Article 50 of the EU AI Act requires providers to ensure that AI systems designed to interact directly with people are built so that “the natural persons concerned are informed that they are interacting with an AI system”, unless that is obvious to a reasonably well-informed person. The Act’s timetable applies Article 50 from 2 August 2026.
California’s companion chatbot law
California’s SB 243, written by state senator Steve Padilla and signed by Governor Gavin Newsom in October 2025, took effect on 1 January 2026. A Fenwick analysis on JD Supra summarises its duties: disclose that the user is talking to AI where a reasonable person could be misled, run protocols for signs of self-harm, add safeguards for minors such as break reminders, and report annually to the state’s Office of Suicide Prevention. It also created a private right of action. Signing it, Newsom said it would protect people from “machines masquerading as human”.
Adam’s Law moves the duty earlier
On 10 September 2026, Newsom signed SB 1119, known as Adam’s Law. It requires covered operators to run a risk assessment before launching any chatbot feature aimed at, or reasonably likely to reach, minors. Penalties reach $5,000 per affected child for each negligent violation and $15,000 per affected child for each intentional one. On those figures, a single intentional violation affecting 1,000 children could cost $15 million (1,000 multiplied by $15,000).
| Rule | In force | Core duty | Enforcement |
|---|---|---|---|
| EU AI Act, Article 50 | 2 August 2026 | Tell people they are interacting with AI unless obvious | National market surveillance authorities |
| California SB 243 | 1 January 2026 | Disclosure, self-harm protocols, protections for minors, annual reports | Private right of action |
| California SB 1119 (Adam’s Law) | Signed 10 September 2026 | Risk assessment before launching features likely to reach minors | Up to $5,000 or $15,000 per affected child, plus private action |
California is not alone. Fenwick notes that other states have passed chatbot-specific laws, most taking effect after 1 January 2027, and that wiretapping claims against chatbots that pass conversations to third-party vendors have grown from roughly two federal cases in 2021 to more than 58 active matters in early 2026.
Designing Chatbots Without Exploiting Anthropomorphism
For any organisation putting a chatbot in front of customers or staff, the research points to a few concrete rules. They are our recommendations, drawn from the evidence above.
Disclose early and plainly
Say it is AI in the first message, not in the terms of service. Article 50 and SB 243 now require this in many settings, and the second-turn finding suggests people’s mental model shifts quickly once a conversation gets going. A clear label at the start anchors expectations before anthropomorphism takes hold.
Do not fake feelings or friendship
Avoid scripted lines such as “I’ve missed you” or “I’m so happy to help you”. They borrow social credit the system has not earned and feed anthropomorphism without adding any service. A warm tone is fine; claims of emotion are not.
Build hand-offs to people
Every chatbot needs a visible route to a human, and anything that might hear distress needs a crisis protocol. Turkle’s lane test is a useful design review: if the bot is drifting into a role a person should fill, it should hand over.
Measure dependence, not just engagement
The MIT and OpenAI trial found worse outcomes among heavier voluntary users. Engagement metrics alone will reward exactly that pattern of deepening anthropomorphism. Track signals such as very long sessions and late-night use, and treat them as a reason to review the design rather than a success. If you are setting rules for AI use internally, our AI strategy and IT governance teams can help.
Anthropomorphism FAQ
What is anthropomorphism in AI?
It is the tendency to treat an AI system as if it had human qualities such as feelings, intentions or understanding. It covers small habits like saying please to a chatbot and larger ones like trusting it as a friend.
Why does anthropomorphism make people say please to chatbots?
Mostly habit. In Future’s December 2024 survey, 82% of polite US users said it was simply nice to be polite; 18% said it was insurance against an AI uprising. Research on real conversations finds politeness grows after the first exchange.
Does being rude to ChatGPT give better answers?
One 2025 study found ChatGPT-4o scored 84.8% with very rude prompts against 80.8% with very polite ones, a difference of 2 questions out of 50. A 2024 study across three languages found rudeness usually hurt. The effect is small and depends on the model.
Is anthropomorphism harmful when a chatbot feels like a friend?
Not always, but heavy use carries risk. A four-week MIT and OpenAI trial with 981 people linked heavier voluntary use to more loneliness and emotional dependence. Children and lonely adults are most exposed.
Do chatbots have to say they are not human?
Increasingly, yes. The EU AI Act’s Article 50 requires it unless it is obvious, from 2 August 2026, and California’s SB 243 has required it for companion chatbots since 1 January 2026.
References
We can’t help treating AI like it’s human. But should we? (TechCrunch)
Artificial Intimacy by Sherry Turkle (Hachette Book Group)
Are you polite to ChatGPT? Here’s where you rank among AI chatbot users (TechRadar)
Exploring Human-LLM Conversations: Mental Models and the Originator of Toxicity (arXiv)
Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy (arXiv)
How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use (arXiv)
Google, chatbot startup to settle teen suicide, self-harm lawsuits (The Hill via MSN)
Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions (Common Sense Media)
Computers are social actors (Wikipedia)
On seeing human: A three-factor theory of anthropomorphism (Psychological Review)
We must build AI for people; not to be a person (Mustafa Suleyman)
Claude Opus 4 and 4.1 can now end a rare subset of conversations (Anthropic)
Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems (EU AI Act)
The Chatbot Compliance Wave: California’s New Law Is Not the Only One to Watch (JD Supra)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
