Hey guys, welcome to TechCrunch’s regular AI newsletter.

This week, a new study in the field of AI shows that generative AI is not all that harmful – at least not in an apocalyptic sense.

In a paper submitted to the annual conference of the Association for Computational Linguistics, researchers from the University of Bath and the University of Darmstadt argue that models like those in Meta’s llama family cannot learn independently or acquire new skills without explicit instructions.

The researchers conducted thousands of experiments to test the ability of different models to perform tasks they had never seen before, such as answering questions about topics outside the scope of their training data. They found that while the models could superficially follow instructions, they were unable to learn new skills on their own.

“Our study shows that the fear that a model will do something completely unexpected, innovative and potentially dangerous is unfounded,” said Harish Tayyar Madabushi, a computer scientist at the University of Bath and co-author of the study, in a statement. “The prevailing narrative that this type of AI is a threat to humanity is preventing the widespread adoption and development of these technologies and is also diverting attention from the real problems that require our attention.”

The study has its limitations. The researchers did not test the latest and most powerful models from vendors like OpenAI and Anthropic, and comparing models is usually an inexact science. But the research is far from the first to find that today’s generative AI technology poses no threat to humanity—and that assuming otherwise may lead to regrettable policymaking.

In an opinion piece in Scientific American last year, AI ethicist Alex Hanna and linguistics professor Emily Bender argued that corporate AI labs are drawing regulatory attention to imaginary, doomsday scenarios to trigger bureaucratic maneuvering. They pointed to OpenAI CEO Sam Altman’s appearance at a May 2023 congressional hearing where he suggested – without evidence – that generative AI tools could “go pretty wrong.”

“The general public and regulators must not fall for this maneuver,” Hanna and Bender wrote. “Instead, we should turn to scientists and activists who conduct peer review and have pushed back the AI ​​hype to understand its harmful effects here and now.”

Her and Madabushi’s arguments are important points to keep in mind as investors continue to pump billions into generative AI and the hype cycle reaches its peak. There’s a lot at stake for the companies backing generative AI technology, and what’s good for them – and their backers – isn’t necessarily good for the rest of us.

Generative AI may not cause our extinction. But it is already causing harm in other ways—think of the proliferation of involuntary deepfake porn, wrongful arrests based on facial recognition, and the hordes of underpaid data annotators. Hopefully policymakers see this too and share this view—or eventually come around to it. If not, humanity may very well have reason to fear.

News

Google Gemini and AI, oh dear: Google’s annual Made By Google hardware event took place on Tuesday, and the company announced a slew of updates to its Gemini assistant — as well as new phones, earbuds, and smartwatches. For the latest news, check out TechCrunch’s roundup.

Copyright lawsuit against AI progresses: A class action lawsuit brought by artists claiming that Stability AI, Runway AI and DeviantArt illegally trained their AIs on copyrighted works can proceed, but only partially, the presiding judge ruled Monday. In a mixed ruling, several of the plaintiffs’ claims were dismissed while others stood, meaning the suit could end up in court.

Problems for X and Grok: X, Elon Musk’s social media platform, has become the target of a series of privacy complaints after it used the data of users in the European Union to train AI models without obtaining users’ consent. X has agreed to stop EU data processing for training Grok – for now.

YouTube tests Gemini brainstorming: YouTube is testing an integration with Gemini to help creators brainstorm video ideas, titles, and thumbnails. The feature is called Brainstorm with Gemini and is currently only available to select creators as part of a small, limited experiment.

GPT-4o from OpenAI does strange things: OpenAI’s GPT-4o is the company’s first model trained on speech as well as text and image data. And that causes it to sometimes behave strangely – for example, mimicking the voice of the person it’s talking to or shouting randomly in the middle of a conversation.

Research paper of the week

There are countless companies offering tools that claim to be able to reliably recognize text written by a generative AI model. This would be useful for fighting misinformation and plagiarism, for example. But when we tested some of them a while ago, the tools rarely worked. And a new study suggests that the situation hasn’t improved much.

Researchers at UPenn developed a dataset and leaderboard, the Robust AI Detector (RAID), with over 10 million AI-generated and human-written recipes, news articles, blog posts, and more to measure the performance of AI text detectors. They found that the detectors they evaluated were “largely useless” (in the researchers’ words), working only when applied to specific use cases and texts that were similar to the text they were trained on.

“If universities or schools were to rely on a narrowly trained detector to detect students’ use of (generative AI) when writing assignments, they could falsely accuse students of cheating when they are not,” Chris Callison-Burch, a professor of computer and information science and co-author of the study, said in a statement. “They could also miss students who cheat by using other (generative AI) to create their homework.”

It seems that there is no magic formula for text recognition using artificial intelligence – the problem is unsolvable.

OpenAI itself has reportedly developed a new text detection tool for its AI models—an improvement over the company’s first attempt—but is declining to release it, fearing it could disproportionately impact non-English-speaking users and be rendered ineffective by minor changes in the text. (Less philanthropically, OpenAI is also concerned about how a built-in AI text detector could affect how its products are perceived—and used.)

Model of the week

Generative AI is apparently not just good for memes. Researchers at MIT are using it to detect problems in complex systems like wind turbines.

A team at MIT’s Computer Science and Artificial Intelligence Lab developed a framework called SigLLM that includes a component for transforming time series data—measurements taken repeatedly over time—into text-based inputs that a generative AI model can process. A user can feed this prepared data into the model and tell it to start detecting anomalies. The model can also be used to predict future time series data points as part of an anomaly detection pipeline.

The framework did not work unusual SigLLM performed well in the researchers’ experiments, but if its performance can be improved, SigLLM could, for example, help technicians identify potential problems in equipment such as heavy machinery before they occur.

“Since this is only the first iteration, we didn’t expect to get it right on the first try, but these results show that there is an opportunity here to leverage (generative AI models) for complex anomaly detection tasks,” Sarah Alnegheimish, a doctoral student in electrical engineering and computer science and lead author of a paper on SigLLM, said in a statement.

Grab bag

OpenAI updated ChatGPT, its AI-powered chatbot platform, to a new base model this month—but didn’t release a changelog (so hardly a changelog).

So what should we make of it? What may a brand of it, exactly? There’s nothing to base it on except anecdotal evidence from subjective testing.

I think Ethan Mollick, a professor at Wharton University who studies AI, innovation and startups, had the right mindset. It’s difficult to write release notes for generative AI models because the models “feel different” from one interaction to the next; they’re largely vibration-based. At the same time, people are using ChatGPT – and paying for it. Don’t they deserve to know what they’re getting into?

It could be that the improvements are incremental, and OpenAI feels it would be unwise to signal this for competitive reasons. Less likely is that the model is somehow related to OpenAI’s reported breakthroughs in reasoning. Regardless, transparency should be a priority in AI. Without it, there can be no trust – and OpenAI has already lost much of it.