Context Window 9
This edition covers Fable’s offensive AI-generated reading summaries and the lessons for publishers, Scott Belsky’s “Small Teams, Big Businesses” 2025 prediction, Google Cloud’s roundup of 300+ generative AI use cases, fresh allegations about Meta’s use of pirated books for training, an Oxford Internet Institute paper on AI’s net positive impact on employment, and a hands-on test of Avataar’s Velocity product-video tool.
Not a good week for book club app Fable, which used AI to generate individual Spotify Wrapped-style summaries for their users, some of which were offensive. Like Apple’s inaccurate summaries of news articles a few weeks ago, it’s a good example of the problem of producing appropriate AI outputs at a scale where full human review is impossible. Fable has responded with safeguards including disclosure that the summary was AI generated, opt-outs, and the ability to flag material as problematic, though those seem insufficient. Of course, it’s easy to be wise in hindsight, and Fable’s internal decision making isn’t clear, but lessons other publishers could learn from this include: *
- Was this scenario considered as a risk in advance?
- Was there specific oversight of training data or was a general model/data used?
- Were system prompts explicitly written to exclude negative and offensive perspectives, as well as offensive vocabulary? What tonal guidelines were given?
- How were AI outputs tested/red teamed?
- Most fundamentally, was editorialising on users’ reading choices ever a good idea?
To paraphrase Jeff Goldblum, just because you can, doesn’t mean you should. I no longer use X, but I was pointed to this thread of 2025 predictions from Adobe’s Scott Belsky. The aspect that’s most interesting for publishing is prediction 5, the emergence of Small Teams, Big Businesses:
“2025 will usher in an era of scaling a business’s reach and aspirations without growing headcount and expenses proportionately… more artisanal-like and privately-owned businesses that may have been uneconomical to run before will start to emerge, powered and made economical by AI-driven tech stacks. Their products and their vision will be intensely personal, but the mechanics of their billing, marketing, etc. will be machine-driven.”
This is undoubtedly a glass half full view, but representative of how many publishers are approaching AI. I know a subscriber to this newsletter building a data publisher on this basis, and many others making a similar differentiation between core product/experience and the processes that deliver them. Thanks to ebooks, capital barriers to publishing are the lowest they have ever been, and AI may have a similar impact in the back office. For inspiration on potential uses of generative AI, this list of over 300 actual use cases from Google Cloud makes interesting reading: it’s light on publishing examples, though there are interesting examples from Spotify and News Corp. Somewhat overshadowed by the other Meta news this week is the allegation that Meta’s decision to use pirated books from LibGen for training purposes was referred up to CEO Mark Zuckerberg. One Meta employee is quoted as writing: “If there is media coverage suggesting we have used a dataset we know to be pirated, such as LibGen, this may undermine our negotiating position with regulators on these issues.”
A new paper from the Oxford Internet Institute argues that AI has had a net positive impact on employment levels, with demand for complementary skills outstripping reduced demand in other areas. As a more hands-on thing to try, I was pointed to Avataar’s new AI tool Velocity this week, which generates product videos from Amazon product page links and it was… not great, but interesting. It produced a fairly banal video in under ten minutes. But it’s currently free, and for zero to one use cases, it might just be enough. I should also caveat this by saying that I didn’t spend time optimising the results, so it’s the equivalent of judging ChatGPT outputs on the basis of a single, unstructured prompt. Really the most interesting aspect is less what the v1 output looks like now, and more how good it gets three or six months from now (and whether a standalone product like this remains viable relative to the major marketplaces building this functionality on a platform level).
This was originally published in my email newsletter. To receive weekly updates on how AI is affecting the publishing industry, sign up here.