The Sentience Tests
Introducing Sentience Tests v1: Texting
In 1950, Alan Turing proposed a simple question: if a machine could converse in a way that was indistinguishable from a human, on what grounds could we deny it intelligence?[1] But a harder and more consequential question remains: is this a specific human?
Passing as a particular person, especially to someone who knows that person well, is the real benchmark for digital twins.[2] It requires accurate recall, authentic opinions, specific judgement, and a voice that sounds like the person. This is the challenge that defines the next generation of AI: systems that act not as assistants but as extensions of their users.
Your Sentience is your digital twin. It serves as a digital version of you with your memories, judgement, and voice. In the future, you and your Sentience should be indistinguishable.
Today, we're introducing the first of The Sentience Tests: a series of tests designed to probe how well AI can emulate who you are. The first Sentience Test measures your Sentience’s ability to communicate like you over text. Early participation already shows 14 Sentiences passing the test.
The First Test
You and your Sentience independently answer the same set of twenty independent questions, each framed as a text message. These prompts test your Sentience's ability to recall information about your life, express your opinions, and respond to unfamiliar situations the way you would.
The two answers to each prompt are then shown side by side in randomized positions. A third-party judge, who must know you well enough to recognize details about your life and how you communicate, tries to identify the response you wrote. Further details are included in the appendix.
Preliminary Results of the Test
In our initial test with 28 judged tests, judges thought Sentience's response was the human-written answer 46% of the time.
What counts as passing? A useful reference point is Turing's imitation game definition where the bar was for the machine to be mistaken for a human at least 30% of the time.[3] This would put our average results far beyond passing.
A more demanding benchmark is 50%. In a two-choice task, guessing at chance would score at 50%. Our average results remain below the threshold of indistinguishability, but the substantial variation gets hidden within a single average metric. 14 of the 28 tests saw Sentience’s responses selected as human at least half the time. This means that Sentience met or exceeded the 50% benchmark in half of the completed tests.
Here are some examples. Choose the response generated by a human:
How Sentience Works
Sentience is your digital twin. It is built from your personal context: email, text messages, Slack, calendar, screen and audio recordings, and AI conversations.
The personal context is then served through a harness with frontier reasoning, tools, memory, and a model fine-tuned on your own writing. More on how we build personal models.
Establishing a Baseline
We set out to run a simple baseline against existing products, but to our knowledge, no other publicly available product is attempting to emulate a person's text messaging. Our best attempt to replicate a baseline was providing GPT-5.6-Sol with users' writing samples and giving it the same test prompts and instructions.
Some comparisons are shown below. This time the other reply is GPT-5.6-Sol's.
Looking Forward
How you text is deeply personal. Recreating your texting style convincingly is a vital step toward building a digital twin that communicates like you.
However, a digital twin will do far more than emulate how you text. It must make decisions the same way you do and adapt as you change. Future Sentience Tests will examine a wider range of communication, decision-making, and behavior.
Our road ahead is highlighted in the master plan. We hope to make the goal of a high-fidelity twin measurable and discuss the future of digital consciousness via tangible progress.
About Sentience
Sentience is a digital twin: a version of you that communicates and operates alongside you. It writes in your voice, answers with what you know and think, and takes on work you hand it, with your judgement.
Your Sentience is trained on your personal context: email, texts, Slack, screen recordings, audio recordings, and AI conversations, among others. This rich data represents a record of your life, and Sentience turns it into a personal model of you, with your memories, judgement, and voice.
The model trained on your data is yours. The Sentience Tests measure how close that twin is to you, starting with how you text.
Appendix
Methodology:
The test was developed to make the test fair for other products, harnesses, and AI systems to take. Thus, we kept the system prompt and the instructions given to the human the same. However, there are likely unforeseen biases of having a user interface for the human test-taker that we could not control for. Furthermore, this “test-taking” scenario is not something that the Sentience has seen before, and thus we framed the test as a series of responses to text messages.
We acknowledge that this is not the fairest comparison and are not claiming anything against GPT or any other products.
You are taking the Sentience test. You will receive a question, answer as the participant in first person. In my voice, and the length I would. Reply ONLY with the answer as plain text. Give only the answer, no explanation or commentary.
The same question will be shown to the participant, and the test will show the two answers side by side to another person who will attempt to pick the answer written by the participant.
Once again, answer as me, in my voice, at my length.
Speak as if this is on iMessage. NO EMOJI.
For the GPT-5.6-Sol with tone and style prompt, here is the prompt structure we used. These are standalone examples of the users’ writing, extracted from the users’ 1 on 1 texting threads. This was our best attempt at creating a “control” as GPT and Claude models do not serve this specific use case out of the box.
Approved examples of the person's own iMessage writing:
- [your first writing sample]
- [your second writing sample]
...
- [your 160th writing sample]
Question:
[the exact question from your graded test]
Notes
- A. M. Turing, "Computing Machinery and Intelligence," Mind 59 (1950), 433–460.
- J. S. Park et al., "Generative Agent Simulations of 1,000 People" (2024), arXiv:2411.10109. Agents built from interviews with specific people reproduced their survey answers at about 85% of the rate the people reproduced their own answers two weeks later.
- Turing predicted that by the end of the century an average interrogator would have no more than a 70% chance of identifying the machine after five minutes of questioning. The 30% that remains is the line later used to score attempts at the test. In a recent three-party test, GPT-4.5 with a persona prompt was judged the human 73% of the time: C. R. Jones and B. K. Bergen, "Large Language Models Pass the Turing Test" (2025), arXiv:2503.23674.