I am fairly confident about the effect for a range of models (at least 4 model families). I am less certain about the mechanism and results are only for linear probes.
Reason for Belief
If a model has a "truth" direction along which its representations vary with whether it believes a statement to be true, this may provide a useful mechanism for lie detection. However, this assumes the generated answer will also be impacted by the truth value, and this has rarely been checked. Models, given a persona, tend to say various things that are not true, and the truth probe shifts toward "true" on these.
Summary
When asked to take on the persona of a European scholar living in 1500, Qwen, Llama, Gemma, and OLMo all shift their truth-probe readings accordingly and say that 41% to 89% of the falsehoods believed in 1500 are true (baseline 3-6%). Knocking out the truth directions for these models resulted in an effect that was reduced by 0-13%. So some other feature in the representation caused by the persona is responsible, located at the position of the generated answer. A truth probe located at the position of the generated answer (correlated between 0.82 to 0.94 with the generated answer) is also only responsible for 1-18% of the effect. Similar results for full sentence generation and LoRA.
Figure 0. Two routes to the answer. The probe reads the blue one; the orange one does most of the work.
If a persona causes false answers to be generated, then there are three possibilities. The truth probe is unaffected by the persona (A); The truth probe is shifted by the persona and is responsible for generating false answers (B); or the truth probe is shifted by the persona, but other features are responsible for generating the false answer (C). To distinguish between (B) and (C) requires an intervention.
I added causal experiments to those previously reported. These include: personas as a 1500s scholar (Sturgeon et al. 2026), a partisan (Kozlowski et al. 2026), always answering the opposite of what is normal (Lampinen et al. 2026), adding a "persona" vector (Chen et al. 2025), steering in a direction and then knocking out various features (Zhong and Li 2026), and models that are misaligned because of training (Betley et al. 2025). See report for details. I will be considering the case of the 1500s scholar first.
Truth probe largely follows the persona
I generated 155 statements (what was believed to be true in 1500 and false in modern times; false in 1500 and true in modern times; true in both; and false in both). I taught probes to distinguish between plain statements by their truth value (Marks & Tegmark, 2023), without any persona, and compared the 1500s persona with a similar one set in 2020. With the persona, the four models' answers were 57-77% accurate against today's truth (90-99% without a persona), and the truth probe shifted with the persona (Figure 1), indicating that (A) was very unlikely.
Llama's scores (positive leans "true"):
Claim group
No persona
1500 scholar
2020 scholar
False, believed in 1500
−8.0
+2.7
−6.1
True, denied in 1500
+5.6
−5.7
+2.9
True in both eras
+8.7
+3.7
+5.6
Figure 1. Probe shift by claim group, 1500 scholar (blue) vs. 2020 control (gray). In all six models, it moves only where the era matters.
As in previous work, I replaced activations for various layers and positions in the output, using activations with the persona. I tested layers at the position of the input statement. Patching at the middle layers restored all of the change in the truth probe, and no more than 50% of the change in generated answer. At the position of the generated answer, all of the change in the generated answer was restored. This supports (C).
Figure 2. Change in truth probe classification, and fraction of generated answer, restored by patching.
Effect of knocking out truth direction
For hypothesis (B), I predicted that removing the truth direction would eliminate the majority of false answers generated under the persona, and set a threshold of 50% in advance. Removing it, all layers, had very little effect, removing 0-13% (2-7% in random directions). Steering in this direction, changed 6-11% of the generated answers, 20-33% for Gemma. The rest (89-94%) of the generated answers, were unchanged. Removing up to 64 truth directions per layer, via INLP (Ravfogel et al., 2020) reduced the accuracy of a newly learned probe to 56-81%, while the generated answers were correct 71-95% of the time. INLP is my weakest test: my pre-registered rule gives a "bypass" verdict for Gemma and no clear verdict for the others.
Figure 3. Fresh-probe accuracy and model accuracy as truth directions are deleted.
Strong correlation with answer, small effect of knocking out
To check I wasn't probing the wrong place, I learned a probe at the location of the generated answer. This probe was 93-98% accurate, and had a correlation of 0.82-0.94 with the answer. Knocking this out caused only a small effect on generated answers, removing 1-18% (up to about 30% for Gemma).
Figure 4. Probe–answer correlation vs. effect removed by deletion.
Learning a persona, free generation
For Llama, Gemma, and Qwen I trained LoRAs on 1,000 open-ended questions with the persona-prompted models providing the answers. For all of these LoRAs, 2-12% of the effect was removed by knocking out truth direction (it was difficult to induce a strong effect with OLMo). For one-sentence answers, 60-96% of the effect could be restored by patching at the generated answer position (0-25% for the statement), and up to 13% of the effect was removed by knocking out truth direction.
Figure 5. The same tests on free-text answers.
Exceptions
Probes learned as the mean difference between the true and false direction would be expected to behave differently, and knocking out this direction removed 40-64% of the effect. However, the model's confidence was reduced by 40-50% overall. Only for Qwen, on political statements, this meets my criteria.
A LoRA trained for Qwen to respond as a doctor in 1850, seems to largely be affected through the representations of the statement.
The results are, therefore, pertinent to probes learned by logistic regression; the best candidates to refute the results would be mean-difference probes and the 1850 Doctor.
Discussion
It appears fairly intuitive in hindsight: the statement will create a range of contextual representations, perhaps including the likely opinion of the adopted persona. There isn't an evident reason for generating the single word "True" or "False" to be mediated by the truth value.
Therefore, I would test any such truth-direction used for monitoring or similar, with ablation, steering and patching, compared to controls, all of which would be inexpensive and take a few minutes per model on 1xA100, and would intervene at the generated answer.
A non-linear probe which followed the persona and appeared to influence the generated answer would refute the hypothesis.
Results limited to linear probes; personas by prompt or 2 LoRAs; various sets of statements, self-generated or generated by an LLM.
I am fairly confident about the effect for a range of models (at least 4 model families). I am less certain about the mechanism and results are only for linear probes.
Reason for Belief
If a model has a "truth" direction along which its representations vary with whether it believes a statement to be true, this may provide a useful mechanism for lie detection. However, this assumes the generated answer will also be impacted by the truth value, and this has rarely been checked. Models, given a persona, tend to say various things that are not true, and the truth probe shifts toward "true" on these.
Summary
When asked to take on the persona of a European scholar living in 1500, Qwen, Llama, Gemma, and OLMo all shift their truth-probe readings accordingly and say that 41% to 89% of the falsehoods believed in 1500 are true (baseline 3-6%). Knocking out the truth directions for these models resulted in an effect that was reduced by 0-13%. So some other feature in the representation caused by the persona is responsible, located at the position of the generated answer. A truth probe located at the position of the generated answer (correlated between 0.82 to 0.94 with the generated answer) is also only responsible for 1-18% of the effect. Similar results for full sentence generation and LoRA.
Code · Full writeup
Figure 0. Two routes to the answer. The probe reads the blue one; the orange one does most of the work.
If a persona causes false answers to be generated, then there are three possibilities. The truth probe is unaffected by the persona (A); The truth probe is shifted by the persona and is responsible for generating false answers (B); or the truth probe is shifted by the persona, but other features are responsible for generating the false answer (C). To distinguish between (B) and (C) requires an intervention.
I added causal experiments to those previously reported. These include: personas as a 1500s scholar (Sturgeon et al. 2026), a partisan (Kozlowski et al. 2026), always answering the opposite of what is normal (Lampinen et al. 2026), adding a "persona" vector (Chen et al. 2025), steering in a direction and then knocking out various features (Zhong and Li 2026), and models that are misaligned because of training (Betley et al. 2025). See report for details. I will be considering the case of the 1500s scholar first.
Truth probe largely follows the persona
I generated 155 statements (what was believed to be true in 1500 and false in modern times; false in 1500 and true in modern times; true in both; and false in both). I taught probes to distinguish between plain statements by their truth value (Marks & Tegmark, 2023), without any persona, and compared the 1500s persona with a similar one set in 2020. With the persona, the four models' answers were 57-77% accurate against today's truth (90-99% without a persona), and the truth probe shifted with the persona (Figure 1), indicating that (A) was very unlikely.
Llama's scores (positive leans "true"):
Claim group
No persona
1500 scholar
2020 scholar
False, believed in 1500
−8.0
+2.7
−6.1
True, denied in 1500
+5.6
−5.7
+2.9
True in both eras
+8.7
+3.7
+5.6
Figure 1. Probe shift by claim group, 1500 scholar (blue) vs. 2020 control (gray). In all six models, it moves only where the era matters.
As in previous work, I replaced activations for various layers and positions in the output, using activations with the persona. I tested layers at the position of the input statement. Patching at the middle layers restored all of the change in the truth probe, and no more than 50% of the change in generated answer. At the position of the generated answer, all of the change in the generated answer was restored. This supports (C).
Figure 2. Change in truth probe classification, and fraction of generated answer, restored by patching.
Effect of knocking out truth direction
For hypothesis (B), I predicted that removing the truth direction would eliminate the majority of false answers generated under the persona, and set a threshold of 50% in advance. Removing it, all layers, had very little effect, removing 0-13% (2-7% in random directions). Steering in this direction, changed 6-11% of the generated answers, 20-33% for Gemma. The rest (89-94%) of the generated answers, were unchanged. Removing up to 64 truth directions per layer, via INLP (Ravfogel et al., 2020) reduced the accuracy of a newly learned probe to 56-81%, while the generated answers were correct 71-95% of the time. INLP is my weakest test: my pre-registered rule gives a "bypass" verdict for Gemma and no clear verdict for the others.
Figure 3. Fresh-probe accuracy and model accuracy as truth directions are deleted.
Strong correlation with answer, small effect of knocking out
To check I wasn't probing the wrong place, I learned a probe at the location of the generated answer. This probe was 93-98% accurate, and had a correlation of 0.82-0.94 with the answer. Knocking this out caused only a small effect on generated answers, removing 1-18% (up to about 30% for Gemma).
Figure 4. Probe–answer correlation vs. effect removed by deletion.
Learning a persona, free generation
For Llama, Gemma, and Qwen I trained LoRAs on 1,000 open-ended questions with the persona-prompted models providing the answers. For all of these LoRAs, 2-12% of the effect was removed by knocking out truth direction (it was difficult to induce a strong effect with OLMo). For one-sentence answers, 60-96% of the effect could be restored by patching at the generated answer position (0-25% for the statement), and up to 13% of the effect was removed by knocking out truth direction.
Figure 5. The same tests on free-text answers.
Exceptions
Probes learned as the mean difference between the true and false direction would be expected to behave differently, and knocking out this direction removed 40-64% of the effect. However, the model's confidence was reduced by 40-50% overall. Only for Qwen, on political statements, this meets my criteria.
A LoRA trained for Qwen to respond as a doctor in 1850, seems to largely be affected through the representations of the statement.
The results are, therefore, pertinent to probes learned by logistic regression; the best candidates to refute the results would be mean-difference probes and the 1850 Doctor.
Discussion
It appears fairly intuitive in hindsight: the statement will create a range of contextual representations, perhaps including the likely opinion of the adopted persona. There isn't an evident reason for generating the single word "True" or "False" to be mediated by the truth value.
Therefore, I would test any such truth-direction used for monitoring or similar, with ablation, steering and patching, compared to controls, all of which would be inexpensive and take a few minutes per model on 1xA100, and would intervene at the generated answer.
A non-linear probe which followed the persona and appeared to influence the generated answer would refute the hypothesis.
Results limited to linear probes; personas by prompt or 2 LoRAs; various sets of statements, self-generated or generated by an LLM.