When Does an A/B Test Stop Being a Test?

Five things worth a second look this week.

01 · SOCIAL + CHILDREN

Thousands of children got a safety button that was not connected to anything

Every app you use is running experiments on you. It is called A/B testing, and it is the ordinary machinery of software: show one group the green button and the other group the blue one, see which gets clicked, keep the winner. Nobody tells you which group you are in, because telling you would ruin the test. This is not a scandal. It is how the thing on your phone got to be the way it is.

New York’s lawsuit against TikTok, reported on 8 October, says the company ran that process on a child safety feature.

The feature is called Algo Refresh. It is meant to let you reset your recommendations, so that if your feed has filled up with something you do not want to see, you can wipe it and start again. The state alleges that TikTok gave some users a version that did nothing at all. They turned it on, they were shown that it was on, and their feeds did not change. Users outside the experiment got the version that worked. Thousands of people were in the test, including teens and children.

TikTok’s answer, given to Reuters, is that this is normal: “we regularly test new products and features to validate the experience and understand how they work in the real world.”

Both of those things can be true at once, which is the whole problem.

Because the allegation is not that TikTok built a feature that failed. A feature that fails is a bug, and every company ships bugs. The allegation is that a working version existed and was withheld, and that the withholding was the point of the exercise. That moves the complaint away from the product and onto the method. It is not “your safety tool was bad”. It is “you were measuring what happens to a child who believes they are protected and is not”.

Which is why this one travels. More than two dozen US states have cases against TikTok along similar lines, and Bloomberg reported in June that another experiment involved withholding a safety feature from around 15 million American users. If a court accepts that running a placebo arm on a protection is deceptive rather than diligent, then the standard practice of the entire industry becomes legally dangerous the moment the thing being tested is a safeguard rather than a shade of blue.

Worth noting what is not established. The detail here comes through Reuters rather than from a document I have read, no court or case number appears in the reporting, and TikTok disputes the characterisation rather than the practice.

Every platform that ships a screen time limit or a content reset is running experiments on somebody. The question this case asks is not whether the dial works. It is whether you were in the group that got the one that does.

02 · TECH + THE OPEN WEB

The open web is closing by text file

Every website keeps a small file called robots.txt. It is a note to search engines saying which pages they are allowed to read. It is not a lock, it is an honour system, and it has quietly governed who gets to see the internet for thirty years.

Reddit is suing several companies for scraping it. One of them, SerpApi, has now counter-sued, and its filing alleges that Reddit serves Google a robots.txt that says yes and serves everybody else one that says no. The split, it claims, began on 1 July 2024. If that is accurate, Bing, DuckDuckGo, Mojeek and Qwant simply cannot index new Reddit conversations, and have not been able to for two years.

The filing puts Reddit’s expanded deal with Google at a reported $60 million a year, announced on the same day Reddit filed to go public. Reddit has not produced the agreement in the case.

So whether a public conversation exists for you now depends on which search box you type into. That is not a philosophical shift in the character of the web. It is two versions of a text file, and a contract nobody outside the court has read.

These are allegations in a live case and Reddit had not responded to them at the time of writing.

03 · AI + COMMERCE

Most of these doors are not being shut. They cannot tell who is knocking.

The pitch for a personal AI assistant is that it does your errands: books the flight, orders the shopping. The problem, reported on 6 October, is that the shops keep throwing it out.

Some of that is deliberate. Amazon has started blocking Meta’s assistant from its retail site. Yelp only admits agents that pay for access through its data licensing programme. Delta says it has no arrangement for third-party agents at all.

But the more interesting cases are the accidents. Walmart’s blocks were not intentional, and it says it actually partners with Meta; the likely culprit was a button asking the visitor to prove they are human. Zillow says it allows agents working for consumers, and that its anti-bot defences catch some of them anyway. United would not say whether it blocks anything, and pointed instead at a clause in its terms forbidding “any robot, spider, other automatic device, or manual process to monitor or copy this website”.

That clause was written to stop scrapers. It is now catching customers.

The web spent twenty years building defences that assume automated traffic is hostile, and those defences cannot tell the difference between a program harvesting prices and a program buying trainers on behalf of the person who owns it. So the fix now being assembled, by Meta, Walmart, Stripe and others, is a standard for sorting good bots from bad.

Which quietly changes the question. Not “is this a robot”, but “whose robot is it, and who says so”.

04 · MEDIA + CREATORS

The number everyone reported is the least interesting one in the survey

Syracuse University’s Center for the Creator Economy asked 1,001 Americans which single source they use most for news. YouTube came first on 16.5%, television news second on 16.1%, print and online media third on 15.7%. Headlines made that a milestone.

It is a rounding error in a survey of a thousand people.

The finding further down is the one worth keeping. Two thirds of the sample said reporting ability is a more important qualification than journalism credentials. Sixty-one per cent said they trust independent news creators as much as network broadcasters.

Distribution moving from one screen to another is an old story and has been for fifteen years. A profession’s licence losing its meaning is not. Training, accreditation and a masthead were the things that told you whose account to believe, and two thirds of people have now decided that the relevant test is simply whether someone can do the work.

One caveat the write-up is honest enough to flag: Syracuse sells creator degrees, so it has an interest in that answer.

05 · SCIENCE + THE INTERNET

The camera was the instrument and the broadcast at the same time

A remotely operated submarine called SuBastian went down into the Mar del Plata Canyon off the coast of Argentina, more than 3,500 metres, roughly twice the depth of the Grand Canyon. It has now been confirmed that it found 57 species and five genera nobody had recorded. Parts of the dive were streamed live and more than four million people watched. The findings went onto an open-access platform rather than into a journal.

So discovery stopped being something announced afterwards by the institution that owned it. People watched it happen.

But the quote in the write-up that stays with you is not about the robot. Dr Daniel Lauretta: “International cooperation in taxonomy is critical, especially in groups with few active specialists.”

Read that again with the rest of it in mind. We can now put a camera three and a half kilometres underwater, hold a live audience of four million while it works, and publish the results to anyone. The thing we are short of is people who can look at what it found and say what it is.

The bottleneck was never the lens.