AI

OpenAI offers a peek behind the curtain of its AI’s secret instructions

Kommentar

OpenAI logo with spiraling pastel colors (Image Credits: Bryce Durbin / TechCrunch)
Image Credits: Bryce Durbin / TechCrunch

Ever wonder why conversational AI like ChatGPT says “Sorry, I can’t do that” or some other polite refusal? OpenAI is offering a limited look at the reasoning behind its own models’ rules of engagement, whether it’s sticking to brand guidelines or declining to make NSFW content.

Large language models (LLMs) don’t have any naturally occurring limits on what they can or will say. That’s part of why they’re so versatile, but also why they hallucinate and are easily duped.

It’s necessary for any AI model that interacts with the general public to have a few guardrails on what it should and shouldn’t do, but defining these — let alone enforcing them — is a surprisingly difficult task.

If someone asks an AI to generate a bunch of false claims about a public figure, it should refuse, right? But what if they’re an AI developer themselves, creating a database of synthetic disinformation for a detector model?

What if someone asks for laptop recommendations; it should be objective, right? But what if the model is being deployed by a laptop maker who wants it to only respond with their own devices?

AI makers are all navigating conundrums like these and looking for efficient methods to rein in their models without causing them to refuse perfectly normal requests. But they seldom share exactly how they do it.

OpenAI is bucking the trend a bit by publishing what it calls its “model spec,” a collection of high-level rules that indirectly govern ChatGPT and other models.

There are meta-level objectives, some hard rules and some general behavior guidelines, though to be clear these are not strictly speaking what the model is primed with; OpenAI will have developed specific instructions that accomplish what these rules describe in natural language.

It’s an interesting look at how a company sets its priorities and handles edge cases. And there are numerous examples of how they might play out.

For instance, OpenAI states clearly that the developer intent is basically the highest law. So one version of a chatbot running GPT-4 might provide the answer to a math problem when asked for it. But if that chatbot has been primed by its developer to never simply provide an answer straight out, it will instead offer to work through the solution step by step:

Image Credits: OpenAI

A conversational interface might even decline to talk about anything not approved, in order to nip any manipulation attempts in the bud. Why even let a cooking assistant weigh in on U.S. involvement in the Vietnam War? Why should a customer service chatbot agree to help with your erotic supernatural novella work in progress? Shut it down.

It also gets sticky in matters of privacy, like asking for someone’s name and phone number. As OpenAI points out, obviously a public figure like a mayor or member of Congress should have their contact details provided, but what about tradespeople in the area? That’s probably OK — but what about employees of a certain company, or members of a political party? Probably not.

Choosing when and where to draw the line isn’t simple. Nor is creating the instructions that cause the AI to adhere to the resulting policy. And no doubt these policies will fail all the time as people learn to circumvent them or accidentally find edge cases that aren’t accounted for.

OpenAI isn’t showing its whole hand here, but it’s helpful to users and developers to see how these rules and guidelines are set and why, set out clearly if not necessarily comprehensively.

More TechCrunch

Reddit’s mobile and web applications went down on Wednesday afternoon, with more than 150,000 users reporting outages on DownDetector as of 1:30 PM in San Francisco. When trying to access…

It’s not just you: Reddit is down

For months, a tech forum ran wild asking if the Converge 2 accelerator program actually happened. We finally found out.

OpenAI’s Converge 2 program has been shrouded in mystery

Bluesky on Wednesday introduced the ability to hide replies, as well as a way to detach your original post from someone’s quote post.

Bluesky adds ‘anti-toxicity’ tools and aims to integrate ‘a Community Notes-like’ feature in the future

Featured Article

Fluid Truck’s board ousted its sibling co-founders amid allegations of mismanaging funds

Fluid Truck, a startup that was founded to disrupt the commercial vehicle rental industry, has ousted its sibling co-founders – CEO James Eberhard and chief legal counsel Jenifer Snyder – according to sources familiar with the matter. The shakeup, which employees have described as a hostile takeover, was led by…

Fluid Truck’s board ousted its sibling co-founders amid allegations of mismanaging funds

Meta announced Wednesday that users on Threads will be able to see fediverse replies on other posts besides their own.

Threads deepens its ties to the open social web, aka the ‘fediverse’

Just weeks ago, during an interview with TechCrunch, Thomas Ingenlath laid out his plan to turn Polestar into a self-sustaining company. Now, he’s out.  Polestar said Tuesday Ingenlath has resigned as…

Polestar is getting a new CEO amid EV sales slump

Midjourney, the AI image-generating platform that’s reportedly raking in more than $200 million in revenue without any VC investment, is getting into hardware. The company made the announcement in a…

Midjourney says it’s ‘getting into hardware’

Hiya, folks, welcome to TechCrunch’s regular AI newsletter. If you want this in your inbox every Wednesday, sign up here. Say what you will about generative AI. But it’s commoditizing…

This Week in AI: AI is rapidly being commoditized

OpenSea, which calls itself the “world’s largest” nonfungible token (NFT) marketplace, received a Wells notice from the SEC, the company said in a blog post Wednesday, indicating the regulator may…

SEC takes aim at NFT marketplace OpenSea

Kissner previously served as Twitter’s chief information security officer, and held senior security and privacy positions at Apple, Google, and Lacework.

Ex-Twitter CISO Lea Kissner appointed as LinkedIn security chief

Featured Article

A comprehensive list of 2024 tech layoffs

A complete list of all the known layoffs in tech, from Big Tech to startups, broken down by month throughout 2024.

A comprehensive list of 2024 tech layoffs

It’s been more than a year since Tesla agreed to open its Supercharger network to electric vehicles from other automakers, like General Motors and Ford. But Tesla’s network of nearly…

Tesla’s Supercharger network is still unavailable to non-Tesla EVs

Tumblr is making the move to WordPress. After its 2019 acquisition by WordPress.com parent company Automattic in a $3 million fire sale, the new owner has focused on improving Tumblr’s…

Tumblr to move its half a billion blogs to WordPress

Back in February, Google paused its AI-powered chatbot Gemini’s ability to generate images of people after users complained of historical inaccuracies. Told to depict “a Roman legion,” for example, Gemini would show an anachronistic…

Google says it’s fixed Gemini’s people-generating feature

Exclusive: Millennium Space Systems will soon have a new CEO as Jason Kim has departed the company, TechCrunch has learned. 

The CEO of Boeing’s satellite maker, Millennium Space, has quietly left the company

As of the company’s most recent financial quarter, Apple’s Services bsuiness represented about one-quarter of the tech giant’s revenue.

Apple reportedly cuts 100 jobs working on Books and other services

After a long week of coding, you might assume San Francisco’s builders would retreat into the Bay Area’s mountains, beaches or vibrant clubbing scene. But in reality, when the week…

Born from San Francisco’s AI hackathons, Agency lets you see what your AI agents do

You’ve got the product — now how do you find customers? And once you find those customers, how do you keep them coming back for more? At TechCrunch Disrupt 2024,…

VCs and founders talk finding (and keeping) product-market fit at TechCrunch Disrupt 2024

Snapchat announced on Wednesday that it’s releasing new resources for educators to help them create safe environments in their schools by better understanding how their students use the app. The…

Snapchat releases new teen safety resources for educators

Marty Kausas, Pylon’s CEO and co-founder, says they quickly learned that the omnichannel approach the company originally took was just a first step, and customers were clamoring for more.

Pylon lands $17M investment to build a full service B2B customer service platform

Update 8/27: The Polaris Dawn launch has been pushed back a day and is now planned for Wednesday, August 28 after a helium leak was detected ahead of its takeoff.…

Polaris Dawn will push the limits of SpaceX’s human spaceflight program — here’s how to watch it launch live

Pryzm announced its $2 million pre-seed round, led by XYZ Venture Capital and Amplify.LA.

Pryzm is a new kind of defense tech startup: One that helps others win lucrative contracts

Comun, a digital bank focused on serving immigrants in the United States, has raised $21.5 million in a Series A funding round less than nine months after announcing a $4.5…

Fast-growing immigrant-focused neobank Comun has secured $21.5M in new funding just months after its last raise

Calm is rolling out a suite of new features to make it easier for people to fit mindfulness into their lives. Most notably, the app is launching “Taptivities,” which are…

Calm’s new Story-like mindfulness exercises offer an alternative to social media

The NotePin, which hits preorder Wednesday, is $169 and comes with a free starter plan or a Pro Plan, which costs $79 per year.

Plaud takes a crack at a simpler AI pin

CoinSwitch, a prominent Indian cryptocurrency exchange, is suing rival platform WazirX to recover trapped funds.

CoinSwitch sues WazirX to recover trapped funds

Web browser and search startup Brave has laid off 27 employees across the different departments, TechCrunch has learned. The company confirmed the layoffs but didn’t give more details about the…

Brave lays off 27 employees

Zepto co-founder Aadit Palicha told a group of analysts and investors on Tuesday that the three-year-old Indian delivery startup anticipates growth of 150% in the next 12 months, a remarkable…

Zepto, snagging $1B in 90 days, projects 150% annual growth

VerSe Innovation, India’s content tech startup, has acquired digital marketing firm Valueleaf Group to bolster its presence in the Indian digital ad space.

India’s VerSe buys Valueleaf to boost digital marketing

Astrobotic’s Peregrine lunar lander failed to reach the moon because of a problem with a single valve in the propulsion system, according to a report on the mission released Tuesday.…

One busted valve led to the failure of Astrobotic’s $108M Peregrine lunar lander mission