Research summaries
What the studies actually say.
Cognition · Stanford HAI · 2026
Fidelity of Style Versus Fidelity of Judgment in Personal Language Models
Chen, Rao, Levesque
Fine-tuning a foundation model on a single individual's corpus produces high stylistic resemblance but only weak agreement with the source individual on novel, unseen decisions.
Security · MIT CSAIL · 2026
Detection Limits of State-of-the-Art Neural Voice Clones
Park, Iyer, Bauer
Open-source detectors trained in 2024 fall to under 60% accuracy on 2026-era commercial voice clones, suggesting a structural disadvantage for post-hoc detection.
HCI · Wharton · 2025
Trust and Productivity in Meetings Attended by Digital Twins
Okafor, Mendes
A field study of 412 knowledge workers using meeting-attendant clones found a 21% reduction in scheduled meeting hours and a small but significant decrease in perceived team trust.
Clinical · King's College London · 2025
Six-Month Outcomes for Bereaved Users of Memorial AI Clones
Hartmann, Liu
A small longitudinal study found that bounded use of memorial clones did not delay grief resolution in most participants, but unbounded daily use correlated with prolonged grief markers.
Economics · NBER · 2026
Labor Market Effects of Personal AI Twins in Knowledge Work
Velasquez, Brennan
Adoption of personal AI twins among salaried knowledge workers is associated with a 9% increase in output and a 3.5% decrease in headcount within firms one year after rollout.
Ethics · Brookings · 2026
Child Safety on Consumer AI Avatar Platforms
Brookings AI Safety Lab
An audit of nine consumer avatar platforms found that six allowed creation of avatars depicting minors without meaningful verification, and four permitted explicit-content prompts on such avatars.