Figure 1
Textual prompt used to obtain ratings of self-distance.
Figure 2
Textual prompt used to obtain ratings of other-distance.
Figure 3
Correlation between LLM self-distance and LIWC self-distance on 100,000 randomly sampled examples from the dataset. Orange points represent the mean LIWC self-distance within one of 100 bins of the LLM self-distance measure. A linear regression line is overlaid in blue.
Figure 4
Effect sizes for mixed-effects regressions, showing relations between internalizing symptoms, linguistic distance, and time in treatment for LIWC self-distance (grey) and LLM self-distance (black) over client text. Effect significance level is indicated in red (i.e., difference between the beta and 0), while significant differences of paired bootstrap comparisons between effects (semipartial ) are shown above the relevant pairs of bars in black asterisks. Ss = Subjects.
Table 1
Client text boundary examples where LLM self-distance High and LIWC self-distance Low maximally disagree.
| LLM SELF-DISTANCE HIGH/LIWC SELF-DISTANCE LOW | ||
|---|---|---|
| LLM SELF-DISTANCE | LIWC SELF-DISTANCE | PARAPHRASED EXAMPLE |
| 7.59 | 0.00 | No matter how the job turns out, it’s just a tool, not a goal. It’s primarily about finances and being able to afford things, rather than being part of my identity. |
| 7.85 | 0.00 | Every now and then, I feel oddly detached, as though I’m witnessing my own life from a distance, fully aware it’s me but feeling as if I’m on the outside looking in. |
| 7.40 | 0.00 | My lawyer and accountant are puzzled by this situation. This process is typically a standard part of any transaction. |
| LLM SELF-DISTANCE LOW/LIWC SELF-DISTANCE HIGH | ||
| LLM SELF-DISTANCE | LIWC SELF-DISTANCE | PARAPHRASED EXAMPLE |
| 1.00 | 10.0 | [-] spent the day at the gym and then joined Mom and her parents for a film shoot! It was such an awesome experience! |
| 1.98 | 10.0 | On Saturday, we slept in, went to yoga, and played cards—it was such a blast! Overall, it turned out to be a fantastic weekend. Haha! |
| 3.10 | 10.0 | We had a fantastic time at a concert tonight! The performances were outstanding, and it was so enjoyable to share the evening together. |
Table 2
Therapist text boundary examples where LLM other-distance and LLM self-distance maximally disagree.
| LLM OTHER-DISTANCE HIGH/LLM SELF-DISTANCE LOW | ||
|---|---|---|
| LLM OTHER-DISTANCE | LLM SELF-DISTANCE | PARAPHRASED EXAMPLE |
| 8.02 | 1.92 | [-] I really want to help you stop feeling stuck and break free from this cloud of unhappiness. The key to your emotional freedom is to grab hold of your specific unhappiness and truly examine it. Say it with me, [-] [-]! If you really want to release emotional suffering, [-], you’ll need to become deeply aware of the emotion that’s causing it—primarily, [-] and loneliness, right? I’d like you to do something for me: step back and just observe your unhappiness. Now, [-], imagine your unhappiness as a wave, coming and going. Picture yourself surfing that wave of unhappiness. Try not to block or suppress it—just observe it. Don’t try to get rid of it or push it away. Don’t hold on to it or amplify it—just surf on it. What was that experience like for you? |
| 7.39 | 1.96 | I sure do, [-]! I really enjoyed getting away—it was so refreshing. Playing it safe is fine, but it can also prevent us from taking chances on things that could be truly amazing in the long run. Weighing options is helpful, but sometimes that voice in the back of our mind keeps saying, ‘What if this, or what if that?’ and before we know it, we start to listen and limit ourselves more and more each day, until we’re living in a small bubble of control… but is that really living? That’s exactly what we’re doing, [-]. We’re challenging those deeply ingrained beliefs and thought patterns. Is it still serving us to think that way, or is it holding us back? Is it worth trying a different mindset and seeing what happens? It could be difficult, and then we could return to safety, or it could open up an entirely new world. |
| 7.04 | 1.53 | Maybe I’m not able to give a clear diagnosis here, and your symptoms don’t immediately bring anything to mind. It might be due to anxiety, depression, or unresolved issues that can crowd your mind when things build up. Picture it like a storage room full of boxes; locked boxes are unresolved issues, and the rest is your living space. Over time, the boxes take up more room. If you’re willing, try this: picture opening one box, observing its contents with curiosity rather than reaction, and emptying it to create more mental space. Let me know if you’d like to try this mindfulness exercise. |
| LLM OTHER-DISTANCE LOW/LLM SELF-DISTANCE HIGH | ||
| LLM OTHER-DISTANCE | LLM SELF-DISTANCE | PARAPHRASED EXAMPLE |
| 2.84 | 7.44 | They seem to have a shared perspective, which both connects them and sometimes leads to tension. |
| 1.91 | 6.93 | The essay about [-] identity… both [-] and [-] are prominent voices in this field and have authored multiple books on the subject. |
| 2.22 | 7.56 | That’s wonderful, and he took it well. He likely appreciated hearing your story and felt grateful that you trusted him with it. |
Figure 5
Correlation between LLM other-distance and LLM self-distance on 100,000 randomly sampled examples from the dataset. Orange points represent the mean LIWC self-distance within one of 100 bins of the LLM self-distance measure. A linear regression line is overlaid in blue.
Figure 6
Effect sizes for mixed-effects regressions, showing relations between internalizing symptoms, linguistic distance, and time in treatment for LLM self-distance (black) and LLM other-distance (grey) over Therapist text. Effect significance level is indicated in red, while results of paired bootstrap comparisons between effects (semipartial ) are shown above the relevant pairs of bars. Ss = Subjects.
Figure 7
Bayesian Mediation models examining whether within-person variance in measures of linguistic distancing mediated changes in internalizing symptoms across time. The 95% CR for the indirect effect did not include zero for any of the models, providing evidence for their significance. Median regression estimates are reported from Bayesian regression models, with their corresponding 95% CRs.
Figure 8
Effect sizes for mixed-effects regressions for the smaller 8-billion parameter model, showing relations between internalizing symptoms, linguistic distance, and time in treatment for LLM self-distance (black) and LLM other-distance (grey) over Client text. Effect significance level is indicated in red. Ss = Subjects.
Figure 9
Effect sizes for mixed-effects regressions for the smaller 8-billion parameter model, showing relations between internalizing symptoms, linguistic distance, and time in treatment for LLM self-distance (black) and LLM other-distance (grey) over Therapist text. Effect significance level is indicated in red. Ss = Subjects.
Figure 10
Pearson’s correlation of LLM self-distance to each of the 66 variables from the full set of LIWC features on client text. A description of what each feature represents can be found in the LIWC manual (pages 5–6): https://www.liwc.net/LIWC2007LanguageManual.pdf.
Figure 7
Bayesian Mediation models examining whether within-person variance in measures of linguistic distancing mediated changes in internalizing symptoms across time. The 95% CR for the indirect effect did not include zero for any of the models, providing evidence for their significance. Median regression estimates are reported from Bayesian regression models, with their corresponding 95% CRs.
Figure 3
Correlation between LLM self-distance and LIWC self-distance on 100,000 randomly sampled examples from the dataset. Orange points represent the mean LIWC self-distance within one of 100 bins of the LLM self-distance measure. A linear regression line is overlaid in blue.
Figure 5
Correlation between LLM other-distance and LLM self-distance on 100,000 randomly sampled examples from the dataset. Orange points represent the mean LIWC self-distance within one of 100 bins of the LLM self-distance measure. A linear regression line is overlaid in blue.