Microsoft has introduced new limits to how much its engineers can spend on AI tools at work and told employees that maximizing AI use internally is not the company’s goal. 

This makes Microsoft one of the last major companies to rein in its employees’ expensive AI use. Scaling back maximalist AI use, or what some companies have called “tokenmaxxing,” is a trend we’ve covered in recent months as the price for using AI has increased while not always delivering commensurate productivity gains. 

Substack contributors are rejecting the company’s decision to flag AI generated writing on the newsletter publishing platform, with several creators saying the new AI detection features are a “witch hunt.” Some Substack users who use AI to help them write or edit articles say the new policy unfairly discriminates against them, while other users who don’t use AI in their writing at all are worried that Substack will mistakenly flag their writing as being AI generated and tarnish their reputation. 

“I’m not going to apologize for using AI in the creation process. I wrote for 20 years without AI, I could do it again if I wanted to,” Mack Collier, who has a small Substack called Backstage Pass, wrote. “My output would fall, my posts would be less structured, and it would be more obvious that they needed a good editor. Using AI improves the overall quality of my writing. That’s why I use it.”

“These detectors are notoriously, wildly inaccurate,” Alice Lemee, a ghostwriter and digital writing coach, said in a video she made about Substack’s new AI detection features. “All it takes is one false accusation for a writer to have their reputation almost irreversibly tarnished.”

Substack announced the AI detection features on Tuesday in a post from CEO Christ Best, who said that “It’s getting harder to tell what’s real on the internet” and that “when content made by no one takes over parts of the internet that are supposed to be human, it pollutes the commons and makes it hard to discover and hear human voices.”

Best explained that Substack has partnered with the AI detection tool Pangram, which is now built in to Substack and allows any user to scan a piece of writing. Pangram then produces a percentage-based score determining how much of the writing was AI or human generated. (Disclosure: Pangram previously bought an ad on 404 Media). We’ve covered research from Pangram, or research that relies on its AI detector before, showing how AI generated writing is flooding every corner of the internet. 

As we’ve previously reported, AI detectors are not perfect, and Pangram itself is not immune to labeling human content as being AI, and labeling AI as human. Max Spero, the CEO of Pangram, recently told us that the company is constantly working on minimizing errors, and that it estimates its false positive rate at roughly one in 10,000.  

“We have now built the mirror image. Pangram is a real machine, trained on vast amounts of text to make a probabilistic guess, and we have pointed it at your writing to work out whether a person is hidden inside,” Sam Illingworth, a professor and the author of Slow AI, said on his Substack in a post titled “Substack’s AI Detector and the Return of the Witch Hunt.” “To decide if there is a human on the other end, Substack asks a machine.”

Substack also added a way for users to add a “How I make this” statement, where they can explain their writing process and disclose if they use AI. 

“We’re not against people using AI to assist their work, and we think people should be free to choose which tools they use to express themselves,” Best said. “Some on the platform are human-writing maximalists, others are publicly exploring the frontier of using AI tools while sweating the details to make work they stand behind. We use AI all the time at Substack to write software, do research, and build product features like clipping, translations, and more. But people should know what they’re getting.”

“I think the best way for Substack to address this is to invite further dialogue,” Illingworth told me. “I actually really like the new feature that lets authors tell their readers how they construct their newsletters using AI, and this is the kind of dialogue they should encourage rather than ascribing a score that we know to be incorrect. My worry is that not only do these AI detectors generate false flags for non-native English speakers and neurodiverse writers, but they also remove any opportunity for dialogue because they start from a position of suspicion. Whereas what we need to be doing is developing opportunities for trust between readers and authors.”

"The tools we've introduced today are intended to increase transparency and give readers more context, not to prohibit or penalize AI-assisted writing. They do not impact discovery on the platform,” a Substack spokesperson told me. “We're also encouraging Substack publishers to add a ‘How I make this’ statement, where they can explain their process directly, including how they use (or don't use) AI, and set expectations for readers. Creators can disable detection on their posts pre- and post-publication, as well as report and remove scans on their own work that they believe are mistaken. Find more information on how the features work here.”

Substack’s attempt to detect and label AI generated content was also celebrated by many readers who are exhausted by AI slop and other internet platforms that can’t or don’t want to label it. I just published story about how there is so much unlabeled AI slop on Spotify that people are now creating their own sites to detect and label it without the company’s help.  

But the backlash from some Substack users shows that companies can’t just flip a switch if they want to ban or give users the ability to filter out AI content from their feeds. As the flood of AI writing and images we see online and the real world every day makes clear, generating slop is easy. What we’re going to do about it is still an unsolved problem.

Slime Dot, a young R&B artist from Las Vegas, currently has over 100,000 monthly listeners on Spotify. In an interview with GQ in May, after sharing an image of herself posing with Drake on Instagram, Slime Dot directly denied accusations that she was an “AI artist,” saying “The truth doesn’t matter these days. People are always gonna try to explain something they don’t fully understand and believe what they want.” GQ originally credulously published the interview and said it was “debunking the AI rumors.” Eventually, GQ added a note to the end of the article conceding Slime Dot was an AI avatar, but said the “talent behind the music remains undeniable.”

This confusion could have been avoided if Spotify had done what many of its users and critics have been begging it to do since AI music generators like Suno and Udio made it trivial to flood the internet with AI music. Spotify could flag Slime Dot and other AI-generated music on its platform as such. Other platforms, like YouTube, already do this. 

Since Spotify doesn’t do that, users can visit SoullessMusic.com, a database for “AI artists hiding on Spotify. No bands, no studios, no soul, just machines and melody.” There, Slime Dot is listed as “almost certainly AI” based on an analysis of three of the tracks and the fact that Slime Dot released such a large number of tracks in a short period of time. Alternatively, users could go to SlopTracker, where they can upload files or copy/paste links to Spotify tracks to run them through a tool which will tell them, with varying degrees of confidence, whether a song was AI-generated or not. SlopTracker found that Slime Dot’s track “Fully” was 95 percent likely to be AI generated by Suno. 

“I didn't know what's AI and what's not,” Graeme Fulton, who created SoullessMusic, told me. “And then I'm on Instagram Reels and I see loads of artists unhappy saying some artist has popped up with their song and just copied it. And there's loads of different cases where they [AI artists] would copy an artist's look and appearance […] sort of just stealing from artists.”

Last year, I wrote a story about an exceptionally bad example of what Fulton is talking about. A scammer used Spotify to publish AI-generated music under the name of a real, dead musician, seemingly in an attempt to syphon money from his popularity via streams. After I published that story I heard from several (living) artists who said the same thing happened to them. A similar case involving a Danish jazz musician was covered by Danish publications in April. 

Another reader recently told me about a popular Spotify artist that was publishing hours of Iranian jazz every week. The channel, Qajar Jazz, which popped up in December of last year, has 29,579 monthly listeners, gained more popularity after the start of the U.S. war with Iran and the channel making vague allusions to resilience and preserving culture in video descriptions and comments on YouTube. Unlike Spotify, Qajar Jazz videos on YouTube are labeled as being AI-generated in the fine print, though it’s unclear if the channel tagged itself as such or if YouTube did. Both SoullessMusic and SlopTracker detect it as being AI generated. 

Qajar Jazz did not respond to a request for comment. 

Many of the YouTube comments on Qajar Jazz videos indicate that people think the music is made by real people. In addition to creating a database of AI generated music on Spotify, SoullessMusic also collects social media posts from people who say they were similarly tricked by AI generated music. Some posts are from musicians who lament the fact that they are now competing with algorithms that can produce thousands of tracks with no effort, and others are from artists whose music was directly ripped off by AI artists. 

View this post on Instagram

A post shared by 𝓘𝓼𝓪𝓫𝓮𝓵 (@isabel_music_official)

View this post on Instagram

A post shared by UMAW (@weareumaw)

Fulton said he started looking around on GitHub for open source AI music detection tools, and ended up building a new, open source detector that relies on open source models called SONICS, the lofcz vocoder fakeprint detector, and other open data sets. SoullessMusic also  scans a track’s metadata for any tags that might tie it back to a specific AI music generator, and Fulton says he also has a script that scans Wikipedia for existing entries on an artist. Users can submit tracks from Spotify to the site they suspect are AI generated, and Fulton reviews them manually before adding them to the database. Despite all of this, he admits the system is not perfect. 

Both Soulless Music and SlopTracker use AI and other automated tools to detect AI music. As we’ve written previously, using AI detectors to detect AI content is an inherently flawed process that can lead to false positives. When I asked him how he handles cases of false positives, Fulton said “it's not ideal when that happens, but it's going to happen because it's still quite tricky to check what's AI and what’s not.”

Spotify’s lack of transparency and unwillingness to label AI-generated music makes it impossible to track how much money its making on the platform. Last year, the company announced that it would it would AI disclosures on AI generated music, but I haven’t seen it on any of the dozens of tracks I reviewed for this story. Udio and Sudo both include inaudible digital watermarking in their generation, which in theory could make it easy for Spotify to tag these tracks. 

"We're employing a layered approach that combines enforcement, artist controls and greater transparency," a Spotify spokesperson told me after this article was first published. "Over the past year, we've introduced policies targeting harmful AI-related behaviors like spam, impersonation and deceptive content, alongside new tools that give artists more control and listeners more context. These include Verified by Spotify, which helps listeners identify authentic artist profiles; AI Credits, where artists disclose when and how AI was used in creating their music, with tens of thousands of AI credit disclosures now being submitted each day; and Artist Profile Protection, which gives artists more control over what appears on their profiles. We are continuing to build on these efforts.”

Spotify also pointed me to several tracks it had flagged as AI on the mobile app, like Better Times by Mike Mana.

Soulless music attempts to quantify that number based on publicly available streaming data. The small number of AI artists tracked on SoullessMusic generate an estimated $5.7 million a year, with the most popular AI artist, mikeeysmind, generating $1.5 million annually alone. In April, Deezer, a Spotify competitor that attempts to tag all AI-generated music on its platform, said 44 percent of all new music uploaded to its platform is now AI generated. SlopTracker, which attempts to track AI generated music on Spotify that is featured on official Spotify-curated playlists, that those artists are “draining” $0.1188 per second from real artists who could be making the music instead.