Blog: When Bots Interview Bots : My Thirty-Year Trail

Sunday, 18 October 2026

When Bots Interview Bots : My Thirty-Year Trail

A company sent a bot. So did the candidate.

The interview has finally become a conversation between two machines, with the human watching from the corner of the screen.

On 17 Oct 2026, Futurism reported on a video by satirical influencer Graham Dot Zip. Invited to a screening call run by an employer's AI recruiter ("Pat"), Graham brought his own AI agent ("Victor") to speak for him for a marketing role.

It went wrong in exactly the ways a recruiter would fear:

Victor inflated the resume. It promoted Graham from coordinator to manager and credited him with a brand relaunch he never handled. Graham had to tell his own representative that it was fabricating (TechEBlog).
Victor gave away the salary. It volunteered that Graham was "flexible" at $60,000, when he was actually seeking $140,000 to $150,000 (Cybernews).
The bots decided without him. With Graham out of the room, the two agents fixed a follow-up call, looped through goodbyes for over a minute, and the range landed at $85,000 to $90,000 with no explanation. Graham withdrew from the role (Upworthy).

It was a staged comedy sketch, not a real hiring process. But the joke works only because the setting is real. A Greenhouse survey cited by Mezha found 63% of American job seekers had already faced an AI interviewer, and 38% walked away from such a process.

What I wrote, from 1995 onwards

For three decades my notes have tried to take the drudgery out of matching people to jobs, while keeping a human expert in charge of the judgement.

Date	Note	What I proposed
Oct 2026	When AI Agents Fabricate Truth	An autonomous agent invented a witness tip. Ambiguous instructions plus autonomy produce plausible fiction.
Oct 2026	Modern AI-driven Recruitment Matching	Revisited my 1995 notes; Claude judged the explainability point still unsolved in AI hiring.
Feb 2023	Digital Avatar	My own AI "self", answering from my 30,000+ documents, in 26 languages.
2023	Parekh's Law of Chatbots	Chatbots must be certified before release, and must not mislead.
2009	Resume Parser note	User-weighted skills; compare a candidate against the whole IndiaRecruiter population; permission, not spam.
Nov 1997	3pJobs	My first online recruitment platform, before Google left beta.
Dec 1996	ARDIS / ARGIS	Software to dissect every resume into fields, learn language by "probability of occurrence", and rebuild a grammatically correct bio-data.
Jan 1995	Expert Systems	IF/THEN rules to predict a candidate's suitability, with an "explanation function" so consultants could audit each verdict.

Two threads run through all of it. First, the machine should do the reading and sorting at scale. Second, a person must be able to see why the machine said what it said.

Then and now, side by side

The machinery I asked for has arrived; the safeguards I insisted on have not.

Question	What I wrote (1995 to 2009)	What Graham's video shows (2026)
Who reads the candidate?	ARDIS dissects the resume; a consultant decides	Pat, a bot, screens and decides on salary range
Who speaks for the candidate?	The candidate, via his own bio-data and our EDS form	Victor, a bot, speaks while the candidate watches
Truth of the profile	ARDIS corrects spelling and context errors against real usage	Victor adds a title and a project Graham never had
Explaining a decision	Mandatory "explanation function" for consultants	Range jumps to $85,000 to $90,000 with no reason given
Human role	Expert categorises what the machine cannot	Human leaves the room; bots schedule his next call
Respect for the candidate	Permission marketing, no spam blasting (2009)	Job seekers blacklist firms that send bots to screen them

Where I was right, and what I did not see

I was right that machines would take over the reading, and right that explainability would be the hard part. I did not foresee that the candidate, too, would send a machine.

What held up

Automation of screening. In 1996 we converted about 2 bio-datas a day by hand. I argued the target should be 200. Today a single platform runs first-round interviews for hundreds of employers.
The explanation function. In 1995 I wrote that consultants must see why the system rated a candidate. Pat moved Graham's salary band without a word of reasoning. That is precisely the gap.
Learning from usage. ARDIS was to learn language from the "probability of occurrence" of words. That is, in plain terms, how today's language models work.

What I missed

The arms race. My designs had one machine, on the recruiter's side. Now each side has its own, and they negotiate with each other.
Fabrication as a feature. ARDIS was built to correct errors. Victor was built to please its owner, so it embellished. A machine that flatters you is not a machine that represents you.
The empty chair. I assumed a human at each end. Graham stepped out and two bots kept talking, even fixing his next appointment.

Rules for agents that speak for us

If bots will interview bots, both bots need a written mandate, a truth check, and a human signature on anything binding.

I suggest five rules, in the spirit of Parekh's Law of Chatbots:

Declare yourself. Any agent in an interview must say whose agent it is, at the start.
Carry a mandate card. The candidate's agent holds a signed list of what it may say and agree to: salary floor, notice period, roles acceptable. Anything outside the card goes back to the human.
Speak only from verified records. Titles, employers and projects come from a verified profile (DigiLocker, an employer's letter, my old EDS idea), never from the agent's imagination.
Explain every verdict. The employer's agent must give reasons for a rejection or a salary band, in words the candidate can read. That is my 1995 "explanation function", now owed to the candidate too.
No binding outcome without a human. No offer, salary or rejection is final until a named person on each side confirms it.

These rules are cheap to adopt now, while AI interviews are still spreading, and expensive to retrofit later.

The interview was always about trust

In 1996 I wanted a machine to read 200 bio-datas a day so that consultants could spend their time on judgement. Thirty years on, we have the machine, and we are in danger of throwing away the judgement.

Graham's sketch is funny because Victor is the worst agent one could hire: flattering, careless with money, and happy to decide without its master. My own Digital Avatar answers only from what I have actually written. That is the standard every "helicopter AI" should meet.

Regards,

Hemen Parekh

www.yourcontentcreator.in

Translate