OpenAI is asking third-party contractors to add actual assignments and duties from their present or earlier workplaces in order that it may well use the info to judge the efficiency of its next-generation AI models, in keeping with information from OpenAI and the coaching knowledge firm Handshake AI obtained by WIRED.
The undertaking seems to be a part of OpenAI’s efforts to ascertain a human baseline for various duties that may then be in contrast with AI fashions. In September, the corporate launched a brand new evaluation course of to measure the efficiency of its AI fashions in opposition to human professionals throughout quite a lot of industries. OpenAI says this can be a key indicator of its progress in the direction of attaining AGI, or an AI system that outperforms people at most economically beneficial duties.
“We’ve employed people throughout occupations to assist gather real-world duties modeled off these you’ve performed in your full-time jobs, so we are able to measure how nicely AI fashions carry out on these duties,” reads one confidential doc from OpenAI. “Take present items of long-term or complicated work (hours or days+) that you simply’ve performed in your occupation and switch every right into a activity.”
OpenAI is asking contractors to explain duties they’ve performed of their present job or previously and to add actual examples of labor they did, in keeping with an OpenAI presentation concerning the undertaking considered by WIRED. Every of the examples needs to be “a concrete output (not a abstract of the file, however the precise file), e.g., Phrase doc, PDF, Powerpoint, Excel, picture, repo,” the presentation notes. OpenAI says folks can even share fabricated work examples created to reveal how they might realistically reply in particular situations.
OpenAI and Handshake AI declined to remark.
Actual-world duties have two parts, in keeping with the OpenAI presentation. There’s the duty request (what an individual’s supervisor or colleague advised them to do) and the duty deliverable (the precise work they produced in response to that request). The corporate emphasizes a number of instances in directions that the examples contractors share ought to replicate “actual, on-the-job work” that the individual has “truly performed.”
One instance within the OpenAI presentation outlines a activity from a “Senior Life-style Supervisor at a luxurious concierge firm for ultra-high-net-worth people.” The purpose is to “Put together a brief, 2-page PDF draft of a 7-day yacht journey overview to the Bahamas for a household who will probably be touring there for the primary time.” It consists of extra particulars relating to the household’s pursuits and what the itinerary ought to seem like. The “skilled human deliverable” then exhibits what the contractor on this case would add: an actual Bahamas itinerary created for a shopper.
OpenAI instructs the contractors to delete company mental property and personally identifiable info from the work information they add. Beneath a bit labeled “Essential reminders,” OpenAI tells the employees to “Take away or anonymize any: private info, proprietary or confidential knowledge, materials nonpublic info (e.g., inner technique, unreleased product particulars).”
One of many information considered by WIRED doc mentions an ChatGPT device known as “Superstar Scrubbing” that gives recommendation on tips on how to delete confidential info.
Evan Brown, an mental property lawyer with Neal & McDevitt, tells WIRED that AI labs that obtain confidential info from contractors at this scale might be topic to commerce secret misappropriation claims. Contractors who provide paperwork from their earlier workplaces to an AI firm, even scrubbed, might be vulnerable to violating their earlier employers’ non-disclosure agreements, or exposing commerce secrets and techniques.
“The AI lab is placing quite a lot of belief in its contractors to resolve what’s and isn’t confidential,” says Brown. “In the event that they do let one thing slip via, are the AI labs actually taking the time to find out what’s and isn’t a commerce secret? It appears to me that the AI lab is placing itself at nice danger.”

