I've been using Readwise Reader as my daily-driver feed reader for just over six months at this point. I still need to write up my overall thoughts on the product, but the headline summary would be: really solid, would largely recommend. I'd say that it's fundamentally changed my workflow and made keeping up with various web feeds a whole lot more enjoyable.
But back in November, there wouldn't have been a need for the caveated "largely" before the "recommend"; it would have instead read "would definitely recommend". So what changed?
In the December feature release, Readwise announced – 🎺 with quite some fanfare 🎺 – a concerning product decision. Whilst Reader has long had "AI" integrations (powered by OpenAI's GPT 3.5 at the moment), they've been fairly easy to ignore. In fact, I'd go so far as to say that the service was a bit of a poster-child for "AI done well". Simple, well thought-through functionality, that was fully opt-in, and which I could see the core user base actually finding beneficial. Their main AI-driven feature was "Ghostreader", which provided quick auto-summarisation of saved documents; for a service designed to streamline knowledge acquisition, this makes a lot of sense.
But in that fateful release, they'd removed one of the core tenants of "good AI": user consent. Rather than being opt-in, their auto-summarisation feature was now universally enabled. Every document, every page, for every user was now being run through a GPT model and spat out as a condensed, sanitised summary. The Readwise team clearly saw this as a beneficial automation – and I don't doubt that people who regularly use Ghostreader had explicitly asked for a way to streamline that process – but there are also some very obvious issues:
- Reader is, at heart, a way to subscribe to and consume other people's content. Content that the reader (either human or software) does not own. So unilaterally deciding that this content can be submitted to LLMs (and the questionable companies that run them) is problematic. This is not, really, your call as an audience member[1], and I'd argue draws parallels with taking video recording equipment into a cinema.
- More specifically, many authors and websites have explicitly blocked the web scrapers and spiders that AI services use to ingest web content for training data. But if a third-party tool – one that authors have not yet blocked[2] – begins firing everything off to OpenAI anyway, which we know results in the query content being churned into the mix as training data, then you're effectively sticking a middle-finger up to the very people you supposedly most admire and want to support.
- On top of which, I (and, I'd guess, most other Readwise users) are simply not going to use summarisation for most of the content we save to the service. Yet despite that, every article is now being run through an LLM request. Requests which are computationally intensive and have an extremely high resource use as a result. It's estimated that every single ChatGPT request uses about 100-1000x the amount of electricity compared to a traditional search engine, and as few as five queries to the service can slurp up a whole bottle of water. Even if that isn't quite hitting the incredibly destructive highs of image generation, that is an obscenely wasteful feature, especially if the output isn't even used!
I therefore imagine I wasn't the only person who immediately submitted a support request to turn this feature off. In fact, based on the response I got back from Readwise's support team (who are pretty excellent, it must be said), it seems they were caught a little off-guard by the volume of such requests. Now that might mean a dozen or a couple of hundred, I have no way of knowing, but I can't help but feel that, actually, this was quite predictable. AI is the latest backfiring hype train, after all.
Thankfully, a little less than a month later, I had a follow-up email from the same support agent: the company had listened. If you want to disable auto-summarisation, you can now do so 🥳
Disabling Auto-Summarisation
Whilst this is a pretty simple process, it's a little hidden away, so I felt it was worth stepping through:
- Log in to Readwise Reader (in a browser; to my knowledge, auto-summarisation never made it to the apps, so neither have these settings);
- Navigate to your Integrations page (click this link or go to Account Settings → Integrations);
- Expand the OpenAI integration by clicking on the "API Key" text with the down-arrow icon;
- You can now toggle auto-summarisation off.
Here's a screenshot of what that looks like to help explain:
Whilst I absolutely applaud the rapidity of getting this "fix" out to users, it does feel a little purposefully tucked away. Personally, I hope that in the future Readwise make the decision to disable this feature by default, as well as move the toggle for auto-summarisation into the account Preferences panel, which feels like a much more logical placement. After all, the other Integrations are user-driven, but I've never asked for or wanted OpenAI to be integrated with my feed reader, so why would I think to look here for these kinds of settings 🤷♀️
Of course, I can think of a few reasons why it might be this hidden away, but they pretty much all require me to view Readwise as a company fundamentally at odds with my own moral compass, and I have to say that other lines of evidence don't back that up. As I mentioned above, their customer service is really quite solid and quick to take things on board; the company as a whole is taking a sensible, long-term approach to product building; and their overall mission is something I fully agree with. Reader is also a piece of software that I've been wanting a company to make for years, and I don't think a software team would end up making it without at least some overlap in terms of world view. Plus, Reader is still very much in beta, so I'm more willing than normal to forgive half-baked features and bad product decisions, particularly when fixes are churned out this quickly.
Still, this is an overall misstep in my books. So much so that, for the last couple of weeks, I have been reviewing the market for possible alternatives. Whilst I think Reader remains leagues ahead of the game (and am very happy that I can now put this issue behind me), going all-in on AI is a red flag that has downgraded my support from "enthusiastic cheerleader" to "wary consumer". I hope that fear is quashed down-the-line, so that by the time Reader enters a full V1 release, I'm back to shouting its praises from the rooftops. I guess time will tell.
