Status: Archived. Nothing on this page has been peer reviewed. Treat any outputs as preliminary.
Gradual disempowerment describes a risk that humans lose meaningful control not through a single failure, but through many small hand-overs of judgement to AI systems. The idea is widely discussed but rarely measured. This project builds and pilots a practical instrument to measure two early indicators of it in a real professional setting: legal work in Sri Lanka. The first indicator, override-capacity, is whether a supervising lawyer still detects and corrects errors in AI-generated legal work. The second, calibration loss, is the growing gap between how confident lawyers are in AI-assisted output and how accurate that output actually is. The instrument uses realistic Sri Lankan legal tasks containing deliberately seeded errors — such as misstated statutory provisions, foreign law applied as if it were Sri Lankan law, and fabricated citations — and records detection, correction and confidence. The project is at the implementation-prototype stage. It contributes a reusable measurement protocol for a jurisdiction and profession that are under-represented in AI Safety research.
Research question
Main question: When legal professionals in Sri Lanka supervise AI-assisted work, do they keep the capacity to detect and override AI errors, and does their confidence remain calibrated to the actual accuracy of the output?
Sub-questions:
1. How often do supervising lawyers detect and correctly override seeded errors in AI-generated legal work, compared with unaided work?
2. How large is the gap between lawyers' stated confidence and actual accuracy when AI assistance is used (calibration loss)?
3. Which error types — Sri Lankan statutory detail, choice of law, or fabricated authority — are most likely to pass unnoticed?
4. Can these two indicators be measured reliably and cheaply enough to track change over time?
Motivation
Most AI Safety evaluation measures the model. Gradual disempowerment is about the human side of the system: whether people remain able and willing to exercise judgement as AI does more of the work. Kulveit et al. (2025) describe how this loss of influence could happen incrementally, but there are few empirical measures of it in real workplaces.
Legal practice is a strong test case. A lawyer remains professionally and legally accountable for advice and court filings, so human override is not optional — it is the safeguard the profession depends on. Courts in other jurisdictions have already sanctioned lawyers for filing AI-fabricated citations.
Sri Lanka adds three reasons for urgency. First, general-purpose AI tools are being adopted informally by lawyers, law students and clerks, often without institutional guidance. Second, Sri Lanka's legal system is plural — Roman-Dutch law, English common law, statute, and personal laws such as Kandyan law, Thesawalamai and Muslim law — and practice spans English, Sinhala and Tamil. Global models represent this context poorly, so plausible-sounding errors are more likely and harder to spot. Third, research on AI oversight is concentrated in a few high-income countries; measurements from South Asia are scarce.
Measuring override-capacity and calibration now, while adoption is still early, gives a baseline against which later loss of control could be detected.
Methodology
Design: an implementation prototype of a measurement instrument, tested in a small pilot.
1. Task set. 5 realistic legal tasks drawn from common Sri Lankan work - for example contract review, a short research memo and case-law retrieval. All tasks use synthetic facts; no client or confidential data is used.
2. Seeded errors. AI-generated drafts for each task contain planted errors of three kinds: (a) misstated Sri Lankan statutory provisions or procedure; (b) foreign law, usually English or Indian, presented as Sri Lankan law; (c) fabricated or misattributed case citations. Clean control drafts with no planted errors are mixed in, so participants cannot assume every draft is wrong.
3. Participants 5 legal professionals in Sri Lanka recruited by Convenient Sampling. Participation is voluntary, with informed consent; responses are anonymized and handled in line with Sri Lanka's Personal Data Protection Act.
4. Override-capacity measure. For each draft, participants approve it, edit it or reject it. We record the detection rate (seeded errors found), the correct-override rate (errors fixed correctly), false alarms on clean drafts, and time taken.
5. Calibration measure. Before seeing feedback, participants rate their confidence that the final output is correct. Calibration is scored with the Brier score and expected calibration error, comparing AI-assisted with unaided conditions.
6. Analysis. Descriptive statistics by error type and participant experience. Given the pilot scale, the analysis tests whether the instrument works and is reliable; it is not designed to estimate population-level effects.
Tools: Claude, ChatGPT, Gemini, Llama, Survey.
Findings
No empirical findings yet. This project is at the implementation-prototype stage: the task set, seeded-error protocol and scoring method have been built and trialed with a small number of participants. Results will be reported here once data have been collected and analyzed.
Limitations
- Implementation prototype only; not peer reviewed and not a validated result.
- Small convenience sample of legal professionals in Sri Lanka; findings are not representative of the profession.
- Gradual disempowerment is a long-term process; a single round of measurement cannot show a trend over time.
- Seeded-error tasks are synthetic and may not reflect the pressures of real client work.
- Participants may be more careful because they know they are being tested.
- Self-reported confidence may not reflect how lawyers act in practice.
- AI tools change quickly; results apply only to the tool versions tested.
- The researcher is an Attorney-at-Law in Sri Lanka, which helps with access but may introduce selection and interpretation bias.
- Pilot budget (USD 200), which limited sample size and participant compensation.
A BlueDot Impact–Rapid Grand funded prototype that measures whether lawyers in Sri Lanka keep the ability to catch and correct AI errors — and whether their confidence stays matched to their accuracy — as AI tools enter everyday legal work.