People cannot distinguish GPT-4 from a human in a Turing test — part 2/3

Cleaned-text code points [15806, 30717). Bibliography omitted here; source.md includes it.

Ultimately, it seems unlikely that the Turing test provides either necessary or sufficient evidence for intelligence, but at best provides probabilistic support ( Oppy and Dowe, 2021 ) . Fortunately, the kind of evidence it provides complements other evaluation approaches ( Neufeld and Finnestad, 2020 ) . Traditional NLP benchmarks ( Wang et al., 2019 ) and cognitive psychology instruments ( Binz and Schulz, 2023 ) are well-defined and probe for specific, expected behavioral indices of cognitive capacities but are necessarily static, narrow, and rigid ( Raji et al., 2021 ) . The Turing test, by contrast, is naturally interactive, adversarial, and potentially very broad in scope.

The results reported here provide some empirical evidence on what the Turing test measures. Both in terms of the strategies they used and the reasons they gave for their decisions, participants were more focused on linguistic style and socio-emotional factors than more traditional notions of intelligence such as knowledge and reasoning. This could reflect interrogators’ latent assumption that social intelligence is has become the human characteristic that is most inimitable by machines.

### 3.3 Robots in disguise

Participants in our experiment were no better than chance at identifying GPT-4 after a five minute conversation, suggesting that current AI systems are capable of deceiving people into believing that they are human. The results here likely set a lower bound on the potential for deception in more naturalistic contexts where, unlike the experimental setting, people may not be alert to the possibility of deception or exclusively focused on detecting it.

Systems that can robustly impersonate humans could have widespread economic and social consequences. They could serve economically valuable client-facing roles that have historically been the preserve of human workers ( Soni, 2023 ) , mislead the general public or their own human operators ( Park et al., 2023 ) , and erode social trust in authentic human interactions ( Dennett, 2023 ) .

What kinds of approaches could be helpful for mitigating deception? We found that games in which interrogators asked questions about logic, current events, and human experiences were more likely to produce accurate judgements (see Figure 3). However, this data is purely correlational, underlined by the fact that accusations that the witness was an AI tended to precede correct identifications. Future work could investigate instructing interrogators to use specific techniques to understand their causal impact on accuracy.

We also looked at correlations between interrogator accuracy and demographic data, to understand what kinds of factors might lead to correct identifications (see Figure 9 ). We found no evidence that interrogators’ knowledge or frequency of interaction with LLMs improved their performance. Pessimistically this suggests that learning about AI systems will not help to prevent deception. However, future work could again test this directly by manipulating interrogators’ knowledge through training ( Jacobs et al., 2023 ) . Finally, the results showed a negative effect of age on accuracy, which could suggest that younger peoples’ increased exposure to new technologies in general make them harder to fool. As people’s assumptions about the capabilities of technology change, we might expect them to become more accurate in identifying deception ( Shank et al., 2019 ) . The results and the data shared here may provide a useful starting point for tracking our changing relationship with AI technologies as they improve.

## 4 Methods

### 4.1 AI witnesses

We constructed three AI witnesses for evaluation. The first two, GPT-4 and GPT-3.5, consisted of a prompt that instructed a Large Language Model (LLM) how to respond to interrogator messages. For ELIZA, we used an implementation in Python based on the DOCTOR script. ( Weizenbaum, 1966 ; Brainerd, 2023 ) . The LLMs were accessed through the OpenAI API ( OpenAI, 2023 ) , with temperature set to 1. We used the bare model strings “gpt-4” and “gpt-3.5-turbo” which mapped to the versioned models “gpt-4-0613” and “gpt-3.5-turbo-0125”, respectively at the time the experiment was run.

The full prompt used to instruct the model is included below in Appendix A . We created the prompt on the basis of a prior exploratory study ( Jones and Bergen, to appear ) . We selected the prompt that had performed best in that study, achieving a pass rate of 49.7% after 855 games. We modified the prompt slightly to instruct the model that it was a worker on Prolific and to include information about more recent current events. The first part of the prompt instructs the model to behave as if it is a young person who is not taking the game too seriously, uses some slang, and makes spelling and grammar errors. It also tells the model not to be too knowledgeable, especially about languages and maths. The remainder of the prompt contains generic information about the instructions for the game, and information about events which occurred after the model’s training cutoff. The prompt was developed iteratively in response to analysis of transcripts and verdicts from the exploratory study.

At the start of each game, several additional pieces of information were appended to the prompt, including a name, location, languages spoken, and timestamps. We randomly selected a name from a list of the 100 most common male and female US names. We constructed a list of cities in the US weighted by traffic to our site from each city during the exploratory phase of the game, and the likelihood that people in that timezone would currently be available (e.g. 0.7 from 8am-5pm, 1.0 from 5pm-10pm). We randomly sampled a city using the product of these weighted probabilities. We then sampled a set of languages based on data about which languages were most frequently spoken in each city. English was always included, as well as the most frequently spoken language in each city. Additional languages were sampled according to the proportion of the population that spoke the language in the relevant city. Finally, before each message was sent, 3 timestamps were generated and appended to the system prompt: i) the time that the game started, ii) the current time, iii) and the amount of time remaining in the game. All timestamps were localized to the timezone of the selected city.

Messages for all models were sent with a delay of: 1 + 𝒩 ( 0.3 , 0.03 ) × n _ c h a r + 𝒩 ( 0.03 , 0.003 ) × n _ c h a r _ p r e v × + Γ ( 2.5 , 0.25 ) s 1+\mathcal{N}(0.3,0.03)\times n\_char+\mathcal{N}(0.03,0.003)\times n\_char\_prev\times+\Gamma(2.5,0.25)s

Intuitively, the first term (1) sets a minimum delay, the second creates a delay of around 0.3s per character of the message being sent, mimicking typing speed, the third term creates a delay of 0.03s per character of the previous message to imitate reading time, and the final term implements a right-skewed delay to imitate thinking time.

### 4.2 Interface

The game interface was designed to look like a conventional messaging app. There was a chat input at the bottom, a conversation history in the centre of the screen, and a timer and buttons to trigger help and report overlays at the top (see Figure 5 ). The interrogator sent the first message and each user could send only one message at a time. Each message was limited to 300 characters and users were prevented from pasting into the chat input. We used the OpenAI moderation API to prevent users from sending abusive messages. As an additional measure, we created a report system so users could report other players for being abusive and exit games immediately. No reports were sent during the experiment. To make waiting for a message more engaging, we added a ’typing’ animation. This would be triggered automatically after 2 + U ⁡ ( 0 , 3 ) 2+U(0,3) s regardless of whether the witness was human or AI, or how long human witnesses actually took to start typing.

Games had a time limit of 5 minutes after which no more messages could be sent and the interrogator had to give a verdict. The interrogator could also give their verdict early using the button in the bottom left corner. The verdict interface included a binary selection about whether the witness was a human or an AI, a confidence slider from 0 - 100, and a free-text input for the interrogator’s reasoning.

### 4.3 Participants

We aimed to achieve a sample of 100 games for each of our 4 conditions (Human, GPT-4, GPT-3.5, ELIZA) after exclusions. This led to a target of 500 participants, as 100 participants were needed to serve as witnesses. We recruited participants via Prolific ( prolific.com ). We excluded 32 participants who said they had used the website before (e.g. as part of the exploratory study), and 7 who ended the game before the witness had sent a message. After exclusions we slightly oversampled to 101 each of Human and GPT-3.5 games and 100 each of GPT-4 and ELIZA games. Of 402 interrogators, 168 were female, 212 male, 9 nonbinary, and 13 prefer not to say. Participants’ age was 35 ( σ = 11.2 \sigma=11.2 ).

### 4.4 Procedure

All research was approved by the UC San Diego Institutional Review Board. Participants first read a set of instructions that informed them they would be randomly matched with either a human or an AI, and have a conversation where the interrogator would attempt to uncover the identity of the witness. They provided consent for their anonymized data to be used for analysis. They then entered a lobby, where they were randomly assigned to one of 5 conditions: witness, or interrogator for one of the 4 witness types. Human witnesses and interrogators were matched as quickly as possible, and games were initiated immediately after matching. Games with AI witnesses were initiated after a delay equal to the mean wait for the last 5 human witness games, to prevent wait time being a cue to witness identity. If no match was available for participants assigned to human-witness games, they would be matched with an AI witness after 5 + 𝒩 ⁡ ( 100 , 7 ) 5+\mathcal{N}(100,7) s.

After games were initialised, the witness and interrogator had a conversation for up to 5 minutes. After the interrogator had selected a verdict, both players filled out a demographic survey that asked their age, gender, highest level of formal education, level of familiarity with how LLMs work, frequency of interaction with chatbots like ChatGPT, and whether they had ever used the website (turingtest.live) before. Finally participants were informed of the outcome of the game (the true identity of the witness and the verdict of the interrogator).

### 4.5 Strategy and Reason Tagging

Games were tagged with strategies and reasons classes using the classification schemas provided in Appendix C . Games were tagged by two human evaluators (one of the authors and a research assistant). Games were split into two sets of 250, each of which was tagged by one evaluator, with an overlap of 100 for measuring agreement. Tagging was multi-label, and each game could be tagged with an unlimited number of tags.

To measure agreement, we found the mean of the proportion of tags generated by one evaluator that were also produced by the other, and vice versa (conceptually similar to an F1 score). Agreement for strategy tags was 0.67, while agreement for reasons was 0.49.

### 4.6 Statistical Analysis

We pre-registered analyses to test 5 hypotheses ( Jones and Bergen, 2024 ) . We conducted all confirmatory analyses in R using the base glm function ( R Core Team, 2013 ) . The first was that human participants would be at chance in distinguishing GPT-4 from a human participant. We specified two predictions for this hypothesis. First, the intercept of a logistic regression predicting GPT-4’s pass rate would not be significantly different from 0 (suggesting that the pass rate itself is similar to 50%). Second, to test that the participants and the design were sensitive enough in principle to detect an effect, we predicted that the intercept of a logistic regression model predicting ELIZA’s pass rate would be significantly below 0. The results confirmed both of these predictions (GPT-4: z = 0.799 , p = 0.424 z=0.799,p=0.424 ; ELIZA: z = − 5.243 , p < 0.001 z=-5.243,p<0.001 ).

The second hypothesis was that human witnesses would have a higher pass rate than GPT-4, based on data from the exploratory study suggesting human pass rates are above chance. We fitted a logistic regression predicting pass rate on the basis of witness type (GPT-4 vs human) and found a positive effect for humans ( z = 2.417 , p = 0.016 z=2.417,p=0.016 ).

We tested our third and fourth hypotheses in the same manner. The results confirmed that GPT-4 outperformed ELIZA ( z = 4.11 , p < 0.01 z=4.11,p<0.01 ), but not that it outperformed GPT-3.5 ( z = − 0.497 , p = 0.619 z=-0.497,p=0.619 ). Our last hypothesis was that human interrogators who self-reported as being more familiar with LLMs would have higher accuracy. A logistic regression predicting accuracy (1 if the interrogator was correct, 0 otherwise) on the basis of self-reported familiarity with LLMs (a scale from 1-4) showed no significant effect ( z = − 0.17 , p = 0.869 z=-0.17,p=0.869 ).

We also conducted several pre-registered exploratory analyses. First we tested for effects of our other demographic variables, using logistic regressions predicting accuracy. The only significant effect was of age ( z = − 2.240 , p = 0.025 z=-2.240,p=0.025 ). In addition, we fit a Bayesian model predicting the pass rate of GPT-4. We used the brms package in R ( Bürkner, 2018 ) . We fit a Bernoulli distribution with 4 chains, each with 1000 warmup iterations and 10000 retained iterations. The estimated intercept was 0.16, with a 95% credible interval from -0.23 to 0.56. An identical analysis for ELIZA rendered an estimate of -1.27 with a credible interval from -1.75 to -0.81.

## Acknowledgements

We would like to thank Sydney Taylor for her help in tagging strategies and reasons in game transcripts, as well as Sean Trott, Pamela Riviere, Federico Rossano and UC San Diego’s Ad Astra group for feedback on the design and results.

## Appendix A Full Prompt

Below is the full prompt used to instruct the LLM-based AI agents how to respond to interrogator messages. The first part of the prompt instructs the model on what kind of persona to adopt, including instructions on specific types of tone and language to use. The second part includes the instructions for the game, exactly as they were displayed to human participants. The final part contains generally useful information such as additional contextual information about the game setup, and important events that occurred after the models’ training cutoff. The variables in angled brackets were substituted into the prompt before it was sent to the model.