Testing: am I part of an experiment?
Am I part of an experiment, and how would I tell?
This chapter explains how influence operations try out narratives before committing to them, and what that looks like from outside: seeded content, trial balloons, and early posts that almost nobody saw. It is candid about how little can be told from a single item, and it shows how to use the early record without feeding the test.
All stages
An honest answer to the stage question
Am I part of an experiment, and how would I tell? Usually you cannot tell from a single item, and it is better to say so than to pretend otherwise. Testing is also ordinary. Advertisers, charities, newsrooms and election campaigns routinely try different wordings on different people and keep what works. What separates that from the covert kind is the same thing that separates white material from grey and black at every stage: whether the tester is honest about who they are.
Testing by a hidden sender is visible, when it is visible at all, as a pattern across many items and usually in hindsight. This stage is therefore less about catching anything in the act and more about reading the early record, and about not volunteering as a data point.
What documented operations did
The clearest public evidence comes from legal process. The 2018 US indictment of the Internet Research Agency describes staff tracking the size of the audiences their pages reached and the engagement each post drew, and reviewing content against those measures. Operations adjust to what they see. Themes that draw a reaction are repeated and others are dropped.
Seeding is the visible edge of this. Secondary Infektion's habit was to place a story first on a small forum or blog, using an account created for the purpose, and then to repost it elsewhere in other languages. The operations OpenAI described in 2024 used language models to turn out short comments and articles in several languages, which makes producing many versions cheap. Neither achieved much. Iteration is evidence of effort and should not be mistaken for evidence of effect.
What early seeding looks like
Looking back from a claim that has become loud, the early record of a seeded narrative tends to share some features. The first appearances are in low-traffic places: comment sections, minor forums, new sites. The accounts that post them have little or no history. Engagement is close to nothing. The same story exists in several versions with different headlines or angles, and later, louder posts cite the early ones as reports.
Two related forms are worth knowing. A trial balloon is an idea floated in a deniable way to see how people react. Politicians do this in plain sight through unattributed briefing, and covert campaigns do it through disposable accounts. A push poll is a survey whose questions carry the message: would you still support this if you knew, followed by a claim. The tell is that nobody appears to be collecting or publishing results.
What to do with a suspected test
The data a test collects is your reaction. Replies, shares, angry quote-posts and poll answers all register as engagement, and mockery counts the same as agreement. Whitney Phillips's guidance for reporters makes the general point: amplifying fringe material, even to condemn it, can give it reach it had not earned. So leave low-engagement material where it lies. Keep dated copies, note the variants, and watch which ones persist.
The early record is also useful. Themes that survive are the ones to prepare for, and the quiet period before they spread is the best time to publish accurate information on them and to prebunk the technique. Keep the innocent reading in view throughout. Most posts that nobody engages with are just unpopular posts, and real people also rephrase an argument until someone answers.
Techniques you will meet here, and the tell
A story is planted in small, low-scrutiny places so that it has a history to point to when it is pushed more widely.
The tell: The earliest appearances are on minor forums or new sites, posted by accounts with no past, weeks before the claim becomes loud.
An idea or accusation is floated deniably to see how an audience reacts before anyone commits to it.
The tell: The claim appears briefly from a source nobody will stand behind, then either vanishes or returns in a firmer form.
Posts are built to draw a reaction of any kind, because the reaction is the measurement.
The tell: The post asks you to comment, vote, tag or correct it, and the content matters less than getting a response.
A survey format is used to deliver a claim and to gauge which version of it stings.
The tell: The questions contain accusations, no organisation is named as running the poll, and no results are ever published.
Many variants of an image or slogan are released, and the ones that spread are reused and refined.
The tell: Near-identical images with small changes of wording appear in quick succession from the same few sources.
Machine-written comments make it cheap to post many versions of the same point and see which draws replies.
The tell: A run of comments makes the same argument in different words, at similar length, with nothing specific to the thread they sit in.
Worked example
Priya goes back through the history of the local groups and the archive of the look-alike site to see how the footbridge themes developed before they became loud. She wants a timeline, and she expects it to prove very little.
- 1She finds posts from seven or eight weeks earlier in small places: the comment section of a regional forum, a buy-and-sell group, a hobby group. They try three angles on the bridge: the cost, safety at night, and incomers on the hill deciding for the harbour. Each drew a handful of reactions.
- 2The look-alike site's earliest articles cover the same three angles, each published under two or three different headlines within days of one another.
- 3In the same period a quick poll for residents went round with no sponsor named. One question asked whether people would still support the bridge if they knew the council had hidden the real cost. That was two weeks before the cost claim spread widely.
- 4From the fourth week the safety angle disappears. Cost and incomers remain and are combined. The incomers angle had drawn the most replies, many from real residents arguing with one another.
- 5She writes the innocent reading beside the timeline: residents also try arguments and drop the ones nobody answers, and this alone would show nothing. What she records as notable is that the variants came from the look-alike site and from accounts with no earlier history.
- 6She lists the two surviving themes as the ones to prepare for, and asks her reporters not to quote-post or mock new variants from the paper's accounts.
Outcome. Priya cannot show that an experiment was run and says so in her notes. She has a dated timeline in which themes narrowed toward whatever drew a reaction, a reasonable expectation of what comes next, and a newsroom rule against supplying engagement to material that has none.
Early record
Fill this in for a claim or theme you are actually tracking. Work backwards from where you first noticed it. Saved in this browser only.
Give the date, the place and the kind of account or site that posted it.
Tick every kind of place where you found early copies.
Choose the closest description.
List the different angles, headlines or wordings you found, with dates.
Note which versions kept appearing and which disappeared.
Tick anything you or your organisation did that would register as a reaction.
Write the ordinary explanation for the pattern, such as real people trying out arguments.
Name the themes most likely to return and the accurate information you would want ready.
Decision check: what does the early record show?
0/7If any of these is not true: Do not publicise the early material and do not announce that a test is under way. Keep collecting dated examples. If you have already amplified a low-engagement item, stop, and do not draw further attention by deleting noisily. If you have no early record at all, say that the origin is unknown and move on to the stages where evidence is easier to gather.
Common mistakes
- ×Amplifying the seed. Exposing or ridiculing an item that a few dozen people saw can show it to thousands.
- ×Over-attribution: reading every failed post or awkward newcomer as a probe. Almost all low-engagement content is ordinary and uninteresting.
- ×Mistaking iteration for effectiveness. An operation that tests carefully can still reach almost no one.
- ×Treating open message testing by a named organisation as sinister. The concern is the hidden sender.
- ×Wasting the quiet period. The time before a theme spreads is when accurate information and a short prebunk cost least and help most.
How do you find out what works without deceiving the people you test on, and who signs off?
Related
References
- US Department of Justice, Special Counsel's Office (2018). United States of America v. Internet Research Agency LLC et al., Indictment (Case 1:18-cr-00032-DLF). US District Court for the District of ColumbiaStaff tracked audience size and engagement for pages and posts and reviewed content against those measures.
- Graphika (Nimmo, B., Francois, C., Eib, C. S., Ronzaud, L., et al.) (2020). Exposing Secondary Infektion. GraphikaStories were first placed on small sites through single-use accounts and then reposted across platforms and languages, with little engagement.
- OpenAI (2024). AI and Covert Influence Operations: Latest TrendsLanguage models were used to produce comments and articles in several languages; the operations drew little authentic engagement.
- Phillips, W. (2018). The Oxygen of Amplification: Better Practices for Reporting on Extremists, Antagonists, and Manipulators Online. Data & SocietyCoverage and condemnation can supply reach to material that had little of its own.
- DISARM Foundation (2022). DISARM Framework (Disinformation Analysis and Risk Management)Preparation and testing of content and narratives as recognised phases before wide distribution.