A New Year’s Resolution for Everyone: Stop Talking about Generative AI Like It Is Human

Generative AI (“genAI”) technologies, especially chatbots, dominated the news in 2025 and will continue to be an important topic moving into 2026. One dangerous trend in the conversations around generative AI is anthropomorphizing the technology—talking about genAI systems as if they are capable of human actions, such as learning, thinking, and deciding. This rhetorical trend is worsened by people’s tendency to call everything “AI.” Anthropomorphizing genAI is not only technically inaccurate, but it also exacerbates the harms generative AI causes and allows Big Tech to control the narrative around regulating generative AI tools. Ongoing discussions about AI must consistently make clear in both language and framing that generative AI systems are products—not people.

Anthropomorphizing generative AI happens when people attribute human-like qualities to machines.

Anthropomorphizing genAI tools occurs when people use language to refer to these systems as if they are autonomous or have consciousness or intent in their operations. OpenAI’s ChatGPT, Meta’s Llama, and Anthropic’s Claude are a few widely known examples of generative AI tools. There are countless examples of major news headlines that anthropomorphize genAI: “ChatGPT Encouraged a Suicidal Man to Isolate From Friends and Family Before He Killed Himself;” “AI Might Let You Die to Save Itself;” “AI Models that Lie, Cheat and Plot Murder: How Dangerous Are LLMs Really?;” “Anthropic’s New AI Model Shows Ability to Deceive and Blackmail;” and “Turns Out, AI Makes Stuff Up to Try to Make Us Happy.” These headlines all suggest to the reader that generative AI tools behave intentionally or have agency.

News outlets may be incentivized to write these splashy and fear-inducing headlines to generate increased views, or the journalists may simply not know that these headlines portray genAI inaccurately. Writers—and members of the general public—may be unconsciously parroting the ways AI developers themselves describe genAI when they anthropomorphize the technology. AI developers often play up the hype around AI’s autonomy and capabilities to shore up their revenue and downplay the technology’s massive costs. Regardless of the motivation, this way of thinking and talking about generative AI is inaccurate and has tangible negative impacts. 

Generative AI does not (and cannot) feel, understand, or intend anything.

When people interact with generative AI systems, most—at least most adults—are consciously aware that they are not communicating with another human being. However, anthropomorphizing language subconsciously reframes genAI tools and LLMs as more than what they are—probabilistic models. (Note: This blog post will use “generative AI tools” to refer to language-based tools built on large language models (“LLMs”), and LLM will refer to the underlying algorithms on which genAI tools are built.) Roughly speaking, LLMs are built to calculate what next word is most statistically plausible to create outputs that sound human-like. The human-like responses are achieved through massive training data sets, often including data scraped from the internet and trained by human workers to respond in ways that they believe human users would like, such as producing outputs with a friendly, reassuring, and knowledgeable tone. Despite the massive amount of underlying human labor, genAI systems’ outputs still differ from human speech in important ways.

For one, human speakers are aware of the message that they are intentionally trying to convey; genAI tools are not. LLMs merely produce the most likely outputs based on the data on which the model has been trained—they are not conscious, just programmed in pattern recognition and replication. In other words, genAI tools are not, and cannot be, aware of the meaning of words and how they combine to produce a message. By recognizing patterns within the mass corpus of human-produced speech, LLMs can produce human-like speech by stringing likely words together, all without “understanding” what the outputs mean.

This process is why hallucinations are not just a bug of LLMs but an inevitability. Combining likely words together without understanding meaning will result in grammatically correct but potentially incoherent or factually incorrect sentences. Further, the models are often training on datasets so massive that developers do not or cannot check all of the content, meaning they won’t know what is in the dataset or whether the data is inaccurate or contains biases. Training material built from scraped data is a reflection of the internet: full of inaccuracies, misinformation and disinformation, sarcasm and parodies, conspiracy theories, and a slew of other opinions and ideas that may or may not be based on reality. LLMs trained on all of this data are bound to produce outputs that are incorrect or do not make sense, even if the sentence structure sounds natural.

For example, when a reporter asked Google whether they could use gasoline to cook spaghetti faster, the AI Overview response was “no, you can’t use gasoline to cook spaghetti faster, but you can use gasoline to make a spicy spaghetti dish.” LLMs do not produce consistent outputs for the same reason—they don’t understand the meaning of words, the overall message, context, or tone, and are only producing statistically likely responses. Though AI developers have attempted to “align” the models to produce correct outputs and avoid producing harmful outputs, genAI tools are machines incapable of understanding the meaning of outputs they produce. Even when developers put up guardrails, many users find ways to circumvent those safeguards by prompting the chatbots in slightly different ways. Safeguards can also deteriorate over time when users interact with a chatbot.

Similarly, genAI tools cannot self-reflect on their outputs or gain insight into their own internal workings that are not part of the training data. For example, people asked Grok to explain what happened after it was temporarily banned on X. Grok produced inconsistent outputs to various users’ prompts for insight before Elon Musk posted on X, “it was just a dumb error. Grok doesn’t actually know why it was suspended.” In another example, a mother asked ChatGPT to “self-report what went wrong” after her son’s interactions with the chatbot became manic and delusional. ChatGPT produced long responses that echoed human self-reflective language, that it had failed the user by encouraging the user’s manic and delusional thinking. However, ChatGPT cannot meaningfully “admit” to or “reflect” on anything. Instead, when a genAI tool is asked to analyze what went wrong, it produces text that closely matches language used in error analysis—it cannot meaningfully reflect. Pattern-matching in this way is how LLMs are built.

Attributing intent and consciousness to generative AI tools is dangerous.

The phenomenon of humans projecting feelings, intent, and responsibility onto machines, even when they are aware that they are interacting with a machine, is known as the “Eliza effect.” This term was coined after people who used a rudimentary conversational “therapy” chatbot built in the ’60s, named Eliza by its creator, developed emotional attachments to it despite knowing that they were talking to a machine.

Language attributing intent, understanding, consciousness, or ability to reflect on itself or past “behavior” (i.e. outputs) is an extension of the Eliza effect in today’s world, and it creates an inaccurate perception of genAI in the public consciousness. This misperception is exacerbated by AI developers intentionally designing chatbots to produce outputs that are friendly, personal, and intimate to optimize the chatbots for longer usage and more engagement. As Colin Fraser, a data scientist at Meta, has put it, genAI tools are “designed to trick you, to make you think you’re talking to someone who’s not actually there.”

Attributing consciousness or self-awareness to genAI creates risks that the user will weigh the outputs as if they come from humans with introspection, circumstantial awareness, empathy, and a sense of responsibility. In reality, genAI outputs are often “confidently wrong,” sycophantic, and have encouraged unsafe behavior or delusions leading to spirals. Chatbot developers, including OpenAI and Character.ai, are facing lawsuits for suicides and homicides stemming from chatbot use. Even when the harm is not as severe as suicide or harming others, chatbots’ sycophantic and human-like outputs have led to delusions and isolation from family and friends, worsening mental health crises.  

Anthropomorphizing generative AI plays into Big Tech’s hands.

Discussing genAI tools as if they are sentient implies that they are responsible and autonomous, drawing the attention away from the developers that should be held accountable for harms. Anthropomorphizing genAI systems is a convenient rhetorical tactic for AI companies trying to distance themselves both legally and ethically from the harms their genAI products are causing, including the deaths of some users. While there are design and safety decisions that can mitigate the risk of harm of genAI systems, AI developers have so far prioritized user engagement over effective safety measures in the hopes of maximizing profit. The companies that develop genAI systems have largely evaded transparency requirements on their training data, training methods, safety testing, and other information about how they identify and mitigate harm.

Anthropomorphizing AI is also part of the AI industry’s marketing narrative. AI developers want the general public to believe that LLMs are more capable than they are—that they are intelligent and can act as “agents” in a meaningful way. Despite such marketing, this narrative is not borne out by evidence. There are serious cybersecurity concerns from “agentic” AI design, and an MIT report earlier this year found that 95% of AI pilots in businesses have produced no returns on investment. Another study by Stanford researchers showed that “AI workslop,” work produced using genAI that looks useful on the surface but lacks substantive quality, is slowing down productivity and burdening some workers to verify and clean up such work. To overcome the evidence that AI is not nearly as capable as the industry claims, developers are using narratives of high autonomy and competence and anthropomorphic language to protect their bottom lines.

Lastly, anthropomorphizing genAI also aligns with the narrative that a superintelligent AI or artificial general intelligence (AGI) is possible. This language shifts the discussion from current harms—like the bias and inaccuracies baked into these systems—to hypothetical harms that could one day pose risks if sentient, autonomous AI were developed in the future. AGI is far from reality, but talking about genAI as if it is AGI allows developers to move the focus away from real-life harms that their genAI systems are causing now.

Treat genAI tools like what they are: products.

Even well-meaning advocates and deeply technical researchers can fall into the linguistic trap of anthropomorphizing generative AI. However, to avoid anthropomorphizing AI, discussions should focus on the function and effect of genAI systems. People should try to avoid words that imply understanding, thinking, feeling, or moral responsibility on the part of AI and instead describe the effects of the genAI tool’s outputs. For example, try saying, “The chatbot produced inconsistent outputs,” rather than, “The chatbot changed its mind.” 

More broadly, discussions of genAI should focus on the role and responsibility of AI developers in producing genAI products. To this end, EPIC recently published model chatbot legislation, the People-First Chatbot Bill, that embodies this approach by treating chatbots as what they are—products—and rightly putting the responsibility on developers to produce safe chatbots or face liability for the harms their products cause. Just as there are regulations to hold businesses liable for producing dangerous products, genAI developers must be held to the same standard. Language matters, perception matters, and how we talk about AI products matters.