presentsUNAIVERSE
The UN AI timeline

    Learn AI /AI Foundations /Anthropomorphism

    Anthropomorphism

    Also called: humanising AI.

    Detail

    In five seconds

    What it is

    Attributing human-like traits, feelings, intentions or understanding to an AI system when its observable output does not establish those inner states.

    In plain words

    People sometimes talk about AI as if it were a person. That can make the system seem smarter, wiser or more caring than it really is.

    An everyday example

    A navigation app says, “Sorry, I made a mistake. I have found a better route.” The wording makes it easy to imagine that the app feels regret, recognises its error and wants to make amends. What can actually be observed is simpler: the software recalculated the route and displayed a conversational message. The apology may make the interface pleasant, but it does not establish remorse, intention or self-awareness.

    Where the story stops working

    A display at a distribution point usually gives a narrow forecast, while many AI systems produce open-ended text, images or dialogue that makes anthropomorphism stronger. Also, a sign does not mimic conversation. Modern chat systems can sustain long exchanges, which makes the human-like impression more persuasive even though the underlying mechanism is still computation, not feelings or lived experience.

    Remember thisHuman-like language is evidence of an output style, not evidence that the system has feelings, intentions, self-awareness or human-like understanding.

    This term is contested

    The general idea is well established, but people disagree about where careful shorthand ends and misleading human-like attribution begins, especially for advanced conversational systems.

    Understand it

    Understand it

    Why it matters

    Anthropomorphism can change decisions. If people treat an AI system like a person, they may trust it too much, rely on it outside its intended use, or fail to check errors. They may also assign blame or authority in the wrong place. For communications and policy work, wording matters: saying a system “thinks” or “knows” can mislead colleagues and the public about its capabilities. Clear language supports accountability, human oversight and safer use.

    What the system does vs what people may read into it

    Help the reader distinguish a model’s actual process from human qualities that people may incorrectly attribute to it.

    1. 01

      Input arrives

      A person enters a prompt, question or other data.

      Top row. Left box: keyboard icon and text ‘Input: question, image, numbers’. Right box: person icon looking at the system.

    2. 02

      Pattern processing

      The model calculates an output from patterns learned during training. It does not need feelings or self-awareness to do this.

      Second row. Left box: network or gear icon and text ‘Pattern matching and computation’. Right box: faded speech bubbles labelled ‘understands me’ and ‘wants to help’.

    3. 03

      Output appears

      The system returns text, a score, a label or another result. If the style sounds human, people may over-interpret it.

      Third row. Left box: screen icon with text ‘Generated answer or score’. Right box: arrows from output to labels ‘intention’, ‘emotion’, ‘belief’.

    4. 04

      Safer description

      Describe the observable function rather than assigning a mind. For example, say ‘the model predicted’ rather than ‘the model knew’.

      Bottom row. Side-by-side sentence cards. Green card: ‘The model generated a reply from patterns in data.’ Red card: ‘The model understood and cared.’

    A two-column comparison. Read each row from left to right. The left side shows the operational step in an AI interaction: input, pattern processing, output. The right side shows the human interpretation people may add, such as intention, emotion or understanding. The final row contrasts a careful description with an anthropomorphic one.

    Hovering or tapping each row highlights the left operational description first, then the right human interpretation, to show that the second is added by the user, not observed directly in the system.

    Describe this without the diagram

    This is a four-row comparison between how an AI system actually works and how people may describe it in human terms. Row one shows that a user gives input such as a question, while the user may already be primed to treat the system like a social partner. Row two shows the model performing computation and pattern matching, while people may infer understanding or intention. Row three shows the output appearing, while people may infer emotion, belief or honesty from tone. Row four contrasts careful wording, such as ‘the model generated a reply from patterns in data’, with anthropomorphic wording, such as ‘the model understood and cared’. The main message is that human-like interpretation is added by people and is not evidence that the system has human mental states.

    With reduced motion: A static two-column table with four rows replaces the highlighting. The left column lists input, pattern processing, output and safer description. The right column lists common human interpretations attached to each stage.

    People often speak about AI as if it were a person. They say a system “knows”, “wants”, “thinks” or “lies”. Anthropomorphism is that habit of giving human qualities to a non-human system.

    What goes in is usually a machine output plus a human reaction. A model receives input data, such as a question or an image. It processes patterns learned from training data, which is the information used to adjust the model during development. What comes out is a prediction, classification or generated text. A person then interprets that output. If the output sounds confident, polite or emotional, people may treat it as evidence of human-like understanding.

    This matters because language shapes judgement. If people describe a chatbot as “understanding the user” or “trying to help”, they may trust it too much, share too much information, or excuse errors that need human review. A convincing style can hide the fact that the system may still produce incorrect, biased or unsafe output.

    Anthropomorphism is not the same as artificial general intelligence, machine consciousness or good user design. A system can sound natural without being conscious. It can respond fluently without having goals or feelings. The safer habit is to describe what the system does in operational terms: it predicts, ranks, classifies or generates output from patterns in data.

    How “The apologetic navigation app” maps onto the real thing

    In the storyIn the AI system
    The display’s estimate about parcel demandA model output based on patterns in data
    The phrase “I think demand will rise”Human-like wording that can invite anthropomorphism
    Staff treating the display as worried or caringPeople attributing feelings or intentions to an AI system
    Past attendance, weather and opening timesInput data used to produce the output

    A worked example

    A desk officer at Headquarters asks an approved drafting tool to review cleared background material before preparing a briefing note for the front office.

    Starting from: Prompt: “Based only on these cleared extracts, identify the delegation’s likely priority and explain your reasoning.”

    1. The model converts the prompt and extracts into numerical representations and generates text from learned patterns and the supplied context.
    2. It replies, “I believe the delegation is mainly concerned about the reporting provision.”
    3. The first-person phrase “I believe” can suggest a considered judgement or settled view inside the system.
    4. In fact, the output is a generated inference. It does not establish belief, intention or an understanding of the delegation’s position.
    5. The desk officer checks the inference against the record, distinguishes explicit statements from interpretation, and phrases any uncertainty accordingly.

    Result: The model may identify a useful pattern, but its first-person language does not make that pattern a belief or diplomatic judgement. Accountability for the briefing note remains with the officer and the clearance chain.

    In a United Nations workplace

    A political affairs officer in an inter-agency secretariat uses an approved tool to suggest codes for non-sensitive, anonymised answers from a needs assessment. The tool writes, “I understand that respondents are frustrated,” and proposes a category. During review, colleagues begin treating that wording as evidence that it has grasped the respondents’ experience and intent.

    Conversational wording can obscure what the tool has actually done: generated a plausible label from patterns in text. The proposed code may still be useful, but apparent empathy or understanding is not evidence that the system has interpreted context as a person would. The officer must review the coding logic and the underlying text.

    Before you try thisNeeds-assessment responses may contain personal or protection-sensitive information. Raw responses must not be entered into a tool that has not been approved for that material and use. Remove unnecessary identifiers, apply the office’s data protection guidance, and retain human review for coding decisions.

    Where you may hear it

    • A vendor demo during an ICT governance board meeting
    • A front-office discussion after an AI-generated briefing note reads unusually confidently
    • An inter-agency working group reviewing conversational tools
    • A procurement evaluation when a chatbot claims that it “understands” users

    Avoid confusion

    Avoid confusion

    Often confused with

    Common misconceptions

    • If a system speaks naturally and responds with empathy, that shows it understands people in a human sense.

      Natural language style can be generated from learned patterns in data. A fluent or empathetic response may be useful, but it is not proof of human-like understanding, feelings or lived experience. The style of the output and the underlying mechanism are different things.

    • Anthropomorphism only matters in science fiction or extreme claims about conscious machines.

      It also appears in ordinary workplace language, such as saying a model ‘knows the answer’ or ‘wants more data’. Those phrases seem harmless, but they can affect trust, risk assessment and responsibility by making the system sound more capable than evidence shows.

    • Avoiding anthropomorphism means never using simple verbs like ‘say’ or ‘answer’.

      Plain language is fine if the limits stay clear. The problem is not every convenient verb. The problem is implying mental states, intentions or awareness without evidence. Saying ‘the chatbot answered’ is often practical; saying ‘the chatbot understood and cared’ goes much further.

    Quick check

    Which statement best avoids anthropomorphism while staying accurate?

    Go deeper

    Go deeper

    The simple version

    People sometimes talk about AI as if it were a person. That can make the system seem smarter, wiser or more caring than it really is.

    The technical version

    Anthropomorphism is the attribution of human-like mental states or traits to an AI system, such as understanding, intention, emotion or personal agency. It often arises from interface design, conversational turn-taking, first-person language, voice, memory cues and coherent text generation. A model may produce outputs that strongly resemble human communication because it has learned patterns from data. That resemblance does not by itself demonstrate subjective experience, self-awareness, stable goals or semantic understanding in the human sense. Machine consciousness is the separate question of whether a machine might have subjective experience. Anthropomorphism concerns human qualities that people infer or project. The key analytical move is to separate observable system behaviour from inferred inner states. It is usually more accurate to describe function, architecture or output behaviour than human-like motives or feelings.

    Reviewed external resources

    Where this came from

    Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

    The one thing to remember

    Human-like language is evidence of an output style, not evidence that the system has feelings, intentions, self-awareness or human-like understanding.