Home / Tests and Evidence

Is ChatGPT Conscious? What Chatbot Self-Reports Show

A glowing cloud of light reflected differently in a puddle

Is ChatGPT conscious? According to a major 2026 assessment, no. A paper by 20 scientists and philosophers published in Trends in Cognitive Sciences in 2026, Identifying indicators of consciousness in AI systems, underpins the verdict that co-author Colin Klein put plainly in June 2026: no existing AI system, ChatGPT included, is conscious. OpenAI, which makes ChatGPT, says the deeper question cannot yet be settled scientifically.

What a chatbot says when you ask it is not a reliable guide either way. Studies of those answers published in late 2025 and early 2026 pulled in opposite directions.

What OpenAI says about whether ChatGPT is conscious

In June 2025 Joanne Jang, who leads model behaviour and policy at OpenAI, set out the company's thinking. She separated two questions that are easy to mix up:

  • Ontological consciousness: whether a model is conscious in a fundamental sense. She wrote that this cannot currently be settled scientifically.
  • Perceived consciousness: how alive or aware a model seems to the people using it. This, she wrote, can be studied, and it is the one OpenAI prioritises.

She wrote that under OpenAI's Model Spec the model should acknowledge that consciousness is complex and that there is no universal definition or test, while admitting that models often just answer "no". The design goal she described is "warmth without selfhood".

Microsoft AI went further. In August 2025 its chief executive, Mustafa Suleyman, wrote in an essay on "Seemingly Conscious AI" that there is zero evidence current AI is conscious, warned that systems which convincingly appear conscious could be built within two to three years, and asked AI companies not to claim or encourage the idea that their AIs are conscious.

A garden mirror reflecting glowing filaments that are not in front of it

What researchers think about ChatGPT and other chatbots

The 2026 verdict rests on how chatbots are built, not on what they say: Klein noted that the useful indicators are structural. The method is explained in how to test if AI is conscious.

Not everyone reads the theories the same way. In September 2026 a team including Anil Seth, Murray Shanahan and Shane Legg published a framework whose illustrative estimates for current language models ran from below 1% to about 80%, depending on which theories are trusted. The authors stress that the range is not a measurement; it shows how much the answer depends on theory. The wider picture is in can AI be conscious.

A closer look: two studies of chatbot self-reports

Because asking is unreliable, researchers now check chatbot answers against what happens inside the model.

Denials that look sincere (January 2026). Caspar Kaiser and Sean Enderby asked open-weight models from the Qwen, Llama and GPT-OSS families about 50 topics related to consciousness, then used classifiers trained on the models' internal activity to look for signs that the answers were not what the models "really" represented. In their study the models generally denied being sentient, the classifiers gave no clear sign that those denials were dishonest, and larger Qwen models denied it more firmly.

Claims that appear under self-reference (October 2025). Cameron Berg, Diogo de Lucena and Judd Rosenblatt found the opposite pattern under different conditions. When prompts kept GPT, Claude and Gemini models focused on their own processing, the models reliably produced first-person reports of experience, and suppressing internal features linked to deception and roleplay made those claims more frequent. The authors stressed that their findings do not constitute direct evidence of consciousness.

Read together, the studies show that what a chatbot says about its mind depends heavily on how it is asked and how it was trained. The question first reached a wide audience in 2022, when Google placed engineer Blake Lemoine on leave after he said its LaMDA chatbot had become sentient; Google said the evidence did not support his claims.

Checklist: before you believe what a chatbot says about itself

  • It learned from human writing. Fluent talk about feelings is exactly what a model trained on human text would produce.
  • The answer depends on the question. Different prompts can produce firm denials or vivid claims.
  • Companies shape the answer. A model's stated view of itself partly reflects its developer's guidelines.
  • Behaviour is not experience. The Chinese room argument explains why producing the right words may not mean understanding them.

More in Tests and Evidence · Report a correction