When ChatGPT pointed, humans verified: how an AI‑surfaced lead reunited a family
On holiday in the south of France, a 66-year-old man typed a few words into ChatGPT and within hours sent an email that led to a phone call and, days later, a face-to-face meeting in London. Decades of dead ends, National Archives visits, radio appeals, even considering a TV help show, had produced nothing. What changed was that an AI model surfaced a public tribute that people had missed for years. The model didn’t “solve” the case; it provided a lead that humans then chased, verified and turned into a reunion.
The lead, as the family tells it
Avtar Singh grew up in Amritsar and says he landed in Halifax, Nova Scotia, on Christmas Eve, 1968. Raised by his paternal grandparents, he carried a single clue from his father: the name Savinder. He spent decades searching, Google, archives at Kew, radio appeals, and came up empty.
“I thought, I’ve tried Google and everything else, why don’t I give AI a shot?” Avtar says. On 8 May (as he recalls), he fed a prompt to ChatGPT while on holiday. According to Avtar, the model returned what he remembers as the “strongest match”:
“The strongest match, based upon your clues, is this … Savinder Kaur Nagi, born in 1936 in Kisumu, Kenya, to Indian Sikh parents.”
That reply pointed him toward an online tribute written by jewellery designer Nicci Dhamu about her mother, Savinder. The piece referenced a 1962 dressmaking certificate from Dhariwal (near Amritsar) and other details that aligned with Avtar’s memories. Avtar says he emailed Nicci at 8.36am that morning; within hours they were on the phone. Nicci, 58, told him the call began with, “Titu, is that you?” and later said, “This is my mum’s son. It’s got to be.”
Nicci: “I wrote it out of grief.”
What followed, and what remains unverified
Nicci’s family moved to the UK from Kenya in 1973. The woman identified in the tribute, Savinder Kaur Nagi, was born in Kisumu in 1936 and died in her sleep in December 2024, aged 88. Nicci says she submitted Savinder’s DNA to Ancestry.com “three or four months before her mother died.” Avtar declined DNA testing, citing privacy concerns and having deactivated Facebook 12 years earlier.
Avtar later ran the same query with Anthropic’s Claude. He reports the model replied: “We don’t want to infringe on anybody’s privacy, so we don’t give out that information.” Avtar also warned the interviewer that “I know that AI can just make up shit. So I thought, OK, that’s nice.”
Both sides describe a rapid and emotional confirmation based on photographs, shared nicknames (Titu/Tito) and family testimony. Both say they do not plan to do DNA confirmation. Important documentary traces, marriage or divorce records in Amritsar, passenger manifests from 1968, or scanned copies of the original dressmaking certificate, are not publicly cited in the family’s account. Where those records exist or do not exist remains an open point outside the family’s testimony.
What this episode actually shows about LLMs and people search
Three realities coexist here:
- LLMs can sometimes surface obscure public fragments, blog posts, niche tributes, community archives, that conventional searches miss by synthesizing the clues a user provides into likely search language.
- LLMs can hallucinate. They do not, in a standard chat, provide auditable provenance for each claim. A returned name or date should be treated as a pointer, not proof.
- The human follow-up is essential. In this case, the model suggested a lead. Humans verified it by contacting the author, comparing photos, and talking to relatives. Those human steps produced the reunion. The model did the pointing. People did the verification and the humane work.
Put another way: AI can shorten the path from hint to contact, but it can’t replace verification or consent.
A practical checklist: how to evaluate an AI‑surfaced lead
- Capture the session. Log the prompt, the exact model response and timestamps. If possible, save a screenshot or export the chat transcript immediately.
- Locate and snapshot the original source. If the model references a webpage or tribute, archive the URL (Wayback Machine) and save a local copy. Don’t rely on the model’s paraphrase alone.
- Corroborate with independent evidence. Look for civil records, passenger manifests, school/employment records, or third‑party reporting before treating the lead as verified.
- Obtain consent before outreach. If you plan to contact a named person based on AI output, document consent workflows and avoid publishing sensitive details without permission.
- Escalate where necessary. If a lead concerns living people or sensitive history, involve legal, privacy, or social‑work professionals before taking action.
Concrete product and policy actions for leaders
CEOs & product teams
- Design the UI to show provenance rules, surface source URLs and a snapshot button whenever a model returns personal data.
- Require human sign-off before outreach, implement a mandatory verification step in workflows that reconnect people.
Privacy officers & legal teams
- Create consent templates and retention policies for AI session logs. Log retention should match legal requirements under GDPR/UK Data Protection law and your internal risk thresholds.
- Assess downstream risks. Personal data aggregation can trigger data‑controller obligations and defamation exposure if uncertain claims are published.
Genealogy teams, social workers, and investigators
- Use AI as triage. Let agents rank leads and surface obscure mentions, but require documentary evidence (certificates, registries, contemporaneous documents) before making reunions public.
- Respect choices around DNA. Biological confirmation is powerful but raises privacy and family dynamics issues. Treat a decision not to test as legitimate and plan alternate verification routes.
Short, honest answers to likely questions
- Did ChatGPT find Avtar’s mother by itself?
ChatGPT (as Avtar recalls) surfaced a public tribute that matched his clues; that suggestion became a lead which humans then followed and verified by contact and family testimony. - Was the relationship confirmed by DNA?
No. Nicci says she submitted her mother’s DNA to Ancestry.com months before Savinder died, but Avtar and Nicci have chosen not to pursue DNA testing to confirm their relationship. - Are LLMs safe for searches about private people?
They can be useful but carry real risks. Models may surface sensitive public content and can hallucinate. Always verify model outputs against primary sources and adopt privacy guardrails. - Should companies building AI agents worry about privacy liability?
Yes. Agents that aggregate or surface personal data should show provenance, require consent before outreach, and include opt-out and human-review mechanisms to manage legal and ethical risks under regimes like GDPR.
There’s something quietly important in how this family handled the lead: they trusted the machine only enough to get a name, then trusted people for everything that mattered, confirmation, conversation, and care. For leaders building AI into workflows, the lesson is simple and operational: let models point, design your product and policy so people verify. That division of labor, models to surface possibilities and humans to verify and to be humane, belongs at the center of any responsible AI strategy.