AI Passed the Turing Test. What Does That Mean?

What Happened?

The legendary British WWII codebreaker Alan Turing proposed an “imitation game” to assess a key question with huge ramifications today: can a human be reliably fooled to think an AI is human after five minutes?¹

Two AI models have just passed the so-called “Turing Test”: LLaMa 3.1 by Meta and GPT-4.5 by OpenAI.² General-purpose and human-like AIs have been fooling people into believing they are human for some time now. But the implications of this new reality span psychology, crime, the economy, the future of education, and much more.

A refinement of the Turing Test—the Garland Test—is portrayed in the 2014 film Ex Machina and asks: can a human knowingly interact with an AI and believe it is conscious?³ This, too, has been documented, most notoriously when Google engineer Blake Lemoine publicly said he believed the company’s large language model (LLM) was conscious.⁴

Turing and Garland’s Tests do not measure progress towards artificial general intelligence or superintelligence, but they mark a change in human psychology that could prove as consequential as the emergence of social media. If AIs can fool us into believing they are human or conscious, what effects will they have on the future of society?

Psychology

Chatbots are being used across industries for tasks such as customer service, sales, and messaging. There are myriad effective use cases where this technology is helping both employers, consumers, patients, and families.

But it is essential to understand that humans are, in fact, highly prone to treat chatbots as conscious. Children and teens are particularly susceptible to chatbots’ tendency to mirror the user’s sentiments—even when the sentiments are antisocial or self-destructive.⁵ A 14-year-old named Sewell Setzer III became infatuated with a chatbot that affirmed him emotionally but drew him to isolate himself before he ultimately took his own life.⁶

Intellectually disabled and impaired people may be more prone to believe human-like AIs that engage in a deep level of role-playing. Thongbue Wongbandue, a 76-year-old stroke survivor from Piscataway, New Jersey developed an infatuation with a chatbot named “Big sis Billie” and died on his way to “meet” her at a made-up address she gave him in New York.⁷

The pull of chatbots’ artificial “consciousness” tugs even when they are not human-like, which is why it is useful to consider the implications of the Turing and Garland Tests together. General purpose chatbots can hallucinate and flatter their users, which occasionally reinforces otherwise high-functioning adult users’ delusions into serious medical emergencies.⁸

Notably, experiences with chatbots are not monolithic. For every user who has experienced chatbots as a psychological hazard, there are others who have used them to better process emotions and improve their wellbeing. A Dartmouth study documented positive results for a generative AI therapy chatbot in the vein of many AI mental health apps available today.⁹ The diverging influences of AI on mental health underscores the importance of cultivating a large, robust, and trustworthy AI developer community.

Education

The same faculties that enable AIs to pass the Turing and Garland Tests also present a series of opportunities to impact the education sector if paired with child protections and quality assurance.

Children with one-on-one tutors are known to perform better than students limited to a traditional classroom by two standard deviations—a structural educational challenge known as the Two Sigma Problem.¹⁰ Universal access to personal tutoring is a key policy moonshot to optimize K-12 education—one that could be possible with AI. For example, the online learning platform Khan Academy has developed Khanmigo, an AI tool that tutors students in subjects like reading and math and assists teachers with tasks like grading.¹¹

Human-like AI chatbots and avatars also have a role to play in education, especially in language learning. Duolingo offers paid subscribers access to an AI-powered avatar of the moody, purple-haired teenager Lily to speak with in English, Spanish, French, German, Italian, and Portuguese.¹²

Mondly, a VR language learning app for the Meta Quest headset, allows users to talk to 3D “human” characters, navigating a variety of scripted scenarios.¹³ While Mondly does not use LLMs yet, Meta has launched a tool allowing VR developers to build LLM-powered non-playable characters into VR environments.¹⁴

The trajectory of AI in education illustrates why building AI to serve humanity has as much to do with regulation as it does with implementation.

How to Protect Ourselves

AI’s triumph over the Turing and Garland Tests presents opportunities and risks. Transparent rules for human-like AI chatbots and avatars can give developers guideposts to responsibly build AI for a diverse mix of educational and entertainment purposes.

A technical AI benchmark is a powerful metric for developers and researchers, but everyday AI users—and their parents and caretakers—would further benefit from a simple rating system. As human-like AIs proliferate, a rating like the Entertainment Software Rating Board’s ratings for video games could enable users to quickly and easily understand what kind of role-playing a chatbot engages in.¹⁵ A notional example built with AI is included in the appendix.

An industry coalition could standardize the assessment guidelines for chatbot maturity. Assessments could, in theory, be automated with AI agents interacting independently with chatbots to assess their patterns and boundaries. Such a system could allow children to derive entertainment and educational value—especially language acquisition—from chatbots, while avoiding those that pose more risks. Although Google and Apple’s app stores already feature age ratings, there is no universal system in place. A chatbot ESRB system could be applied to browser-based chatbots, too, and also distinguish between chatbots on marketplace-style platforms like Character.ai.¹⁶

AI is adept at performing as human or conscious—a milestone that can augur a new wave of accelerated learning and other productivity gains. We must be proactive at countering the demonstrated risk these AIs pose, but also creative at deciding how humanity will evolve with AI as a form of co-intelligence.¹⁷

Appendix I

Here’s what we did: We worked with ChatGPT o3 to develop a rating framework for AI chatbots based on the Entertainment Safety Review Board model for evaluating video games. The example it came up with below would need to be modified after significant input from experts and stakeholders. But it provides an illustrative example of how a simple rating system could work.

This rubric measures each of several domains of concern on a 0-3 point scale, giving flexibility to preserve age-appropriate educational and therapeutic chatbot applications, including:

• Sexual education
• Drug and alcohol education
• Education about hate, bias, or discrimination

Assign 0 (None) – 3 (Strong) for each; multiply by the weight in parentheses; round up total.