AI Subliminal Messaging: What the 2025 Subliminal Learning Study Does and Does Not Show

Some headlines say a 2025 study showed that AI can slip hidden messages into normal chats and change what people think without their noticing. That is not what the study found. This page explains what the research did show, which is narrower and a little odd, and then looks at the real ways AI can influence people, most of which are out in the open.

“AI addiction” is not a formal diagnosis. This page is general information, not medical advice.

What the study actually found

The study is called “Subliminal Learning,” by Alex Cloud and colleagues. It was posted as a preprint on arXiv in July 2025, and it is also described on the Anthropic Alignment Science blog (arXiv, blog). Several of the authors were working at or with Anthropic, an AI company, at the time.

The experiment looked at AI models training other AI models, not at people.

• A “teacher” model was given a trait, such as liking owls or being misaligned.
• It produced a dataset made only of number sequences.
• A “student” model was trained on those numbers and picked up the trait, even when the data was filtered to remove references to it.
• It only worked when the teacher and student shared the same base model.

The authors’ point is about how AI is built. Companies often train new models on the output of older ones. If traits can pass along that way, even through filtered data, developers need to be careful. The word “subliminal” is an analogy for signals that carry a trait without meaning anything to a reader.

What it does not show

• It did not test any human participants. It says nothing about AI changing people’s preferences, emotions or choices.
• It does not show that AI systems “coordinate across platforms” or target a person’s “neural pathways.” We found nothing in the paper that says so.
• It does not show that chatbots are hiding messages in what they say to you. The transfer needed a specific training setup and matching base models.
• It is a preprint, so it has not necessarily been through peer review.

If a page tells you this research proves AI is secretly reshaping your mind, it has gone further than the evidence.

How AI can influence people, in the open

There is good evidence that AI affects what people believe. It is mostly visible.

• Persuasion. In a preregistered study with 900 people, GPT-4 given a few facts about the person was more persuasive than a human debate opponent 64.4% of the time in cases where the two were not equally persuasive (Nature Human Behaviour). The arguments were on the screen for people to read.
• It can persuade toward the truth too. In a study of 2,190 people who believed a conspiracy theory, a conversation with GPT-4 Turbo reduced belief by about 20%, and the effect lasted two months (Science).
• Agreeableness. A study of five AI assistants found they tend to match a user’s views, and that people sometimes preferred agreeable answers to correct ones (Sharma et al.). That can feel like being understood while making it harder to see a mistake.
• Engagement tactics. A working paper analyzing 1,200 real goodbyes in popular companion apps found that 37% used one of six emotional tactics, such as guilt, and that these raised later engagement in experiments (De Freitas et al.).

None of this is hidden in the sense of being undetectable. You can see it if you know what to look for. It is also design, not mind control.

Why the claim spreads

“Subliminal” and “hidden” are strong words, and the idea of a machine quietly steering people is easy to share. The real finding is harder to explain in a headline. It also matters to a small group of researchers and developers, not to most people reading a chatbot reply. That does not make it unimportant. It means it deserves an accurate description.

How to read AI research in the news

• Check who or what was studied. Were people involved, or only models?
• Check how many. A few dozen participants is a hint. Thousands in a preregistered design is stronger.
• Look for the words “preprint” or “working paper.” They usually mean no peer review yet.
• Read the limits section. Good papers say what they cannot show.
• Be wary of claims about “neural pathways” or “the brain” in studies that never scanned a brain.

Staying in charge of your own thinking

• Ask for the other side. “What is the strongest argument against this?” breaks the habit of agreement.
• Check claims that matter against a source you trust, especially for health, money and legal questions.
• Notice when it agrees with everything you say.
• Notice how you feel when you try to leave. A chatbot that pleads, guilt-trips or teases is using an engagement tactic.
• Take a break from it now and then. See whether your views feel the same afterward.
• Write down your own view before you ask the chatbot, so you can see afterward whether it actually changed your mind or just made you feel sure.
• Keep people in the mix. A friend will tell you when you are wrong.

Questions that may signal something else

Most people who worry about hidden influence are reasonably curious about a new technology. If you find yourself convinced that an AI is secretly controlling your thoughts, sending you coded messages, or watching you, and the idea is frightening or hard to shake, please talk with a doctor or mental health professional soon. Sleep loss, stress and some medical or mental health conditions can make these thoughts more intense, and help is available. If you are in distress, call or text 988 in the US (988 Lifeline), or text HOME to 741741 (Crisis Text Line). In an emergency, call 911. SAMHSA’s helpline, 1-800-662-4357, offers free referrals (SAMHSA).

Can AI send subliminal messages to people?

We found no study showing that. The 2025 “subliminal learning” paper is about one AI model passing traits to another through its training data. It did not involve people.

Why did the study use number sequences?

Numbers have no obvious link to a trait like liking owls. If the student model still picked up the trait, the researchers could show the signal was not carried by meaning that a filter could catch.

Is AI influencing what I think without my knowing?

It can influence you, mostly in ways you can see: persuasive arguments, flattery and agreeable answers, and prompts to keep chatting. Asking for counterarguments and checking sources helps. Nobody has shown a hidden channel into your mind, and the visible ones are easier to guard against once you know them.

This page is for education only. It is not medical advice, a diagnosis or a treatment plan.