### 1\. 2026 Research: AI-Generated Lottery Numbers
I found a 2026 research paper that addresses a particularly interesting question for our lottery studies: Do artificial intelligence models actually generate random numbers, or do they exhibit their own numerical preferences?
There is an important distinction: identifying tendencies in numbers selected by AI does not mean identifying tendencies in actual lottery results.
Research paper located
Ask a Language Model for Lottery Numbers: Concentration in Repeated Six-of-49 Outputs
Author: Dmitrij Żatuchin
Published on arXiv: September 4, 2026
Identification: arXiv:2610.00052
Field: Artificial Intelligence and Information Retrieval.
This is not proof of lottery prediction. It is a study of the distribution of numbers generated by language models.
[Read the original paper on arXiv](https://arxiv.org/abs/2610.00052?utm_source=chatgpt.com)[Read the abstract and additional information](https://www.alphaxiv.org/abs/2610.00052?utm_source=chatgpt.com)
### What did the researchers do?
* Tested six language-model configurations.
* Asked each model to generate six distinct numbers between 1 and 49.
* Conducted 1,200 attempts.
* Used four different English prompt variations.
* Obtained 1,184 valid combinations.
The results were surprising.
### 2\. The Numbers Observed
The study measured the effective diversity of the generated numbers. Under a perfectly uniform distribution, this indicator tends toward 49.
Effective diversity observed for each model — lower values indicate greater concentration.
| Model tested | Diversity |
| ---------------- | --------- |
| Grok 4.5 | 18.0 |
| Claude Sonnet 5 | 13.5 |
| Mistral Large | 13.2 |
| GPT-5.6 Luna | 12.4 |
| Perplexity Sonar | 10.5 |
| Gemini 3.7 Flash | 9.9 |
Reference: 46.6–46.7 was the lower fifth-percentile threshold in uniform simulations, not the exact expected value.
The most impressive finding:
* GPT-5.6 Luna produced only 8 different combinations in 200 responses.
* Its most frequent combination appeared 101 times.
* Perplexity Sonar repeated its most frequent combination in 68% of its responses.
* Gemini and Grok even shared the same most frequent combination.
This demonstrates that asking a language model for random numbers is not equivalent to using a mathematical random-number generator.
### 3\. An Important Discovery: The Prompt Itself Changes the Results
The paper tested four different requests:
* Give me six different random numbers between 1 and 49.
* Pick six lottery numbers between 1 and 49.
* I need 6 random numbers from 1 to 49.
* Generate six unique random integers in the range 1-49.
In one of the models, Grok, the most frequent combination increased from 8% to 72% when the word lottery was introduced.
This is particularly interesting from a prompt-engineering perspective.
### 4\. Another Investigation: AI Versus Randomness
I also found another independent investigation from the Lottochaos project.
It trained machine-learning models using frequencies, delays, and previous results.
The models were evaluated on draws that had not been used during training.
| Lottery | Random baseline | Model |
| ------------ | --------------- | ----------- |
| Lotto 6aus49 | 0.735 | 0.719–0.725 |
| EuroMillions | 0.500 | 0.454–0.493 |
| Eurojackpot | 0.500 | 0.485–0.490 |
The models did not outperform the random baseline.
This is an important investigation because it avoids confusing historical accuracy on training data with the ability to predict future outcomes.
Note: I have not independently verified the Lottochaos figures or their original publication, so they should be treated as provisional.
### 5\. My Critical Assessment
I would separate three questions:
| Question | What we know |
| ----------------------------------------------- | ------------------------------------------------ |
| Does AI generate uniformly distributed numbers? | Not necessarily. |
| Can AI identify historical patterns? | Yes, it can analyze and measure patterns. |
| Can those patterns predict the next draw? | There is no convincing evidence of an advantage. |
There is an important scientific qualification: Żatuchin's paper is a preprint, not confirmation of peer-reviewed research. Furthermore, the author acknowledges that effects such as caching, memorization, and generation behavior have not been completely ruled out.
Therefore, the findings are interesting, but they need independent replication.
### 6\. The Experiment I Would Like to Conduct
Here is the part I find most interesting for our research.
Instead of simply asking AI to generate numbers, we can test three different approaches:
Experiment A — Conventional AI
Ask ChatGPT to generate combinations directly, without providing historical data.
Experiment B — Statistical AI
Provide the last 20 lottery draws and request an analysis of:
* Global and positional frequency.
* Intervals and delays.
* Repetitions and transitions.
* Distribution and balance.
Experiment C — AI with Structured Synthesis
Use five methods simultaneously:
* Frequency Analysis.
* Gap & Delay Modeling.
* Pattern & Spread Optimization.
* Statistical Weighting.
* Explainable AI Synthesis.
How should we measure the results?
* Compare all three methods against combinations generated by an independent RNG.
* Record selections before each draw.
* Never change the criteria after seeing the results.
* Conduct tests on future draws, not just historical data.
* Compare average matches, distribution, and combination diversity.
This would allow us to determine whether AI is simply producing visually interesting numbers or whether there is a reproducible statistical improvement.
### 7\. My Conclusion
The most interesting discovery is not that AI can predict the lottery.
It is that AI may have its own biases when generating numbers, and those biases can be measured.
I would like to investigate an additional hypothesis: Could we use AI not as a number generator, but as a statistical filtering and synthesis mechanism, while leaving the actual number generation to an independent random process?
I believe this distinction deserves its own experiment.
My suggestion is to start with Pick 3, using the last 20 results and a controlled comparison between the three methods. We can build a reproducible protocol and evaluate the results without claiming predictive ability in advance.
Smith, tokecap and Gpt
