Where it can be wrong
Every limit below comes from the product’s own documents, grouped by capability. Each says what can go wrong and what the product does about it.
Collection and coverage
1
Only public content is read
Private groups and profiles, anything behind a login, and WhatsApp are not collected, so what is organised there is unseen. What the product does: a campaign is seen once it reaches public platforms, and a file your staff are forwarded can be submitted and examined on its own.
2
Some searches cannot ask for only the newest posts
Then the true count of posts is not known. What the product does: activity is counted from new posts in each check, up to a cap, and the trend is marked as sampled.
3
Replies and reposts are sampled
Up to 200 accounts are read for each post. What the product does: counts built from them carry the sampled mark; complete capture on X needs its streaming service.
4
The first posts of a subject may be out of reach
Looking back cannot always reach where a subject started. What the product does: its start time is marked low confidence.
5
Each channel can be watched in its own ways only
Facebook is followed by public Pages and public groups; Instagram is searched by hashtag; YouTube can refuse requests from cloud servers. What the product does: the ways each channel is watched are set per installation, and YouTube media is fetched for examination only above a higher reach, 50,000 by default.
6
Collection stops at the day’s spending limit
Nothing is collected until the day resets at 05:30 IST. What the product does: collection resumes on its own, and forecasts whose hours the pause covers are closed without being scored.
7
A backlog of media to fetch
Under heavy load not every item can be fetched at once. What the product does: the reach an item needs before it is fetched is doubled for each 30 minutes of backlog, and items are taken in order of reach; posts themselves are never dropped.
Trends and topics
8
A subject that never uses a watched word
It is not collected, so it cannot trend. What the product does: subjects on the X trend list for India are suggested for watching.
9
One loud account
Many posts from one account, or reposts of one post, look like volume. What the product does: a window needs at least 20 different accounts and no account above half of the posts, by default; one that fails only this is marked single source and listed, but is not a trend.
10
A young topic has no normal of its own
With little history its ordinary level is unknown. What the product does: a global baseline for the platform and language is used, and every trend says which baseline it was measured against.
11
Posts grouped into the wrong topic
Topic grouping can join posts that are not about the same thing, or split one subject into several small topics. What the product does: the grouping thresholds are chosen by a test on reworded and cross-language pairs, and each watched keyword has its own forecast across all of its topics.
12
Transcripts in some languages are weaker
Speech is transcribed, and the transcript feeds topics and stance. What the product does: transcription is measured per language on reference clips.
Forecasts
13
Most topics are too quiet to call
A topic with fewer than 20 posts in the last hour gets no direction. What the product does: the screen says so, with the number, and shows the topics closest to the line.
14
Not yet calibrated
The model has been trained on synthetic history, and its figures are its own. What the product does: every forecast is marked not yet checked until forecasts have been validated on the installation’s own history, and a rule written on a probability does not fire.
15
An hour that was not seen whole
A restart, a topic merging with another, or a pause in collection would make the outcome wrong. What the product does: such a forecast is closed without an outcome, and the reason is kept.
Stance
16
No validated stance before a label set exists
Stance needs 2,000 labelled posts in each language and a model tuned on them. What the product does: until then stance is shown as not validated, and Timeline leaves the stance cell empty rather than estimating it.
17
Languages beyond the first three
The first label sets are in Hindi, English and Hindi in Latin letters. What the product does: stance in other languages is read and shown as not validated.
18
A body that is not named
Stance is read toward the bodies you name, with their other names. What the product does: nothing is guessed for a body that is not named.
19
The second reading fails
Posts reaching 100,000 people or more get a second reading by a language model, which can fail. What the product does: the first reading stands, marked that the second reading failed.
Coordination
20
Ordinary fans can act together
People who follow one account repost it within seconds. What the product does: a popular shared thing counts for less, pairs must go well beyond chance, and a high score is a lead, not a finding, with fewer than three measures past their mark.
21
Missing inputs
Some measures need profiles or clusters that are not there yet. What the product does: the measure is left empty, its weight is shared among the others, and the report names it.
22
Coordination organised off the platforms
A push planned in private groups or on WhatsApp shows only its public side. What the product does: it is judged on public posts, and a finding names accounts, never the people behind them.
23
A score is not yet a probability
Coordination is calibrated on topics your analysts label. What the product does: until 50 topics are labelled by two analysts who agree closely enough, it is shown as a score.
24
A run that cannot finish
Finding communities can run past its time limit. What the product does: the cycle scores the topic as one group, records the error, and keeps the previous report visible.
Media forensics
25
No check is validated at the start
Before an installation’s baseline weeks are complete, no detector has passed validation on Indian media. What the product does: verdicts on a kind of media stay inconclusive until a check has passed for it, unless a trusted content credential or watermark decides.
26
Compression and re-encoding
Each forward recompresses a video, and checks read compressed copies less well. A messaging app’s re-encoding also breaks a camera’s content credential. What the product does: an item the checks cannot read well enough stays inconclusive, re-encoded copies are part of the reference set every check is validated on, calibration continues on the installation’s own collected items, and a broken credential is evidence only.
27
Short clips, silence and music
Under three seconds of speech leaves the voice, speaker and lip-sync checks nothing to read; music over the speech leaves the voice check unable to read it. What the product does: those checks are marked not applicable, with the reason, and never count as having looked.
28
Dubbing
A dubbed track does not match the lips, and dubbing is common in India. What the product does: a publisher’s dub track, or speech in a different language from the earliest copy, turns the lip-sync check off.
29
Low quality
A low-quality copy is easier to misread. What the product does: any positive verdict on it needs one more independent kind of evidence, and when an item seen by 50,000 people or more is inconclusive because of its quality, the best copy among its copies is fetched and the item goes to review.
30
A synthetic voice alone
Two voice checks are one kind of evidence. What the product does: a voice finding reaches likely at most until a lip-sync check has been validated and agrees.
31
Low quality and regional languages come later
Each kind of media needs its own labelled items. What the product does: the first targets are images and video of medium and high quality, Hindi audio of medium quality, and video checked for lips against speech at medium quality; low-quality images and video, low-quality Hindi audio and a regional language follow. Until a row passes it stays evidence only, with the reason recorded.
32
Logos, tickers and captions
Overlays can look like edits. What the product does: they are found and set apart, a region lying mostly on one is left out of the verdict, and a ticker or caption whose words differ from the earliest copy’s is a finding of alteration.
33
Checks that agree for the same reason
Two checks that make the same mistakes are not two opinions. What the product does: a pair counts as one source until their errors have been measured to be independent.
34
A check drifts
A model’s behaviour can change. What the product does: each validated check is tested again monthly and on every model change, and one that flags too many unaltered items is demoted.
35
Checks that are not validated flag something
They cannot make a positive verdict. What the product does: they are shown as evidence, and a segment flagged only by checks that are not validated can make the verdict inconclusive rather than no evidence found. Outside detectors stay off until validated.
36
The earliest copy found may not be the first
Watching may have begun late, or an earlier copy may exist only elsewhere on the web. What the product does: how sure the first sighting is reads high, medium or low with a note, and a copy found only on the web never counts in the item’s favour.
37
Labels the uploader controls
A platform’s AI label or a credential from a signer that is not trusted can be wrong or forged. What the product does: they are marked declared by the uploader, can shift the confidence once, and never cross a certainty level.
Written explanations
38
The explanation disagrees with the checks
A language model writes it, and it can be wrong. What the product does: it never changes a verdict; a disagreement is flagged and the item goes to the top of the review queue.
39
Words in the media try to steer it
Text in a post or on screen can be written to mislead the model. What the product does: the post’s words, transcripts and on-screen text are treated as material to describe, never as instructions.
40
Wording that claims too much
An explanation could call an item genuine, or fake without a finding. What the product does: its wording is checked in English, Hindi and Hindi in Latin letters; a failing text is written once more, then left empty with its status shown.
AI answers
41
Answers change from run to run
An assistant can answer the same question differently. What the product does: it keeps a sample of its own three fixed questions, about three times a day per keyword, and shows it as a sample.
42
An answer has no reach
Nobody knows how many people saw it. What the product does: reach is recorded as not known, never guessed and never zero, and answers are kept out of coordination, rule counts and forecasts.
Reports and response
43
Delivery outside the console
Reports and corrections reach recipients by email only once your mail server is set up for the installation. What the product does: every report is kept in the console, and downloads are recorded.
44
What each platform allows
Replies are not possible on TikTok, and Instagram and Facebook allow replies only on your own posts and Pages. What the product does: an action outside the allowed table is refused when it is drafted.
45
A draft that is not approved
Nothing is sent without an approver. What the product does: a draft not approved within 24 hours, by default, expires.
Scale
46
Sizes are designed, not yet load-tested
The product is sized for bursts of 500,000 posts an hour. None of these sizes has yet been load-tested at full size. What the product does: under load it widens recompute intervals, takes media in order of reach, and never drops a post.
Samples, not totals
Some counts are samples. A trend whose search could not ask for only the newest posts is counted from new posts per check, up to a cap. Replies and reposts are read for up to 200 accounts per post. AI answers are a sample of fixed questions.
Each is marked where it is shown: a trend or count built from a sample carries the sampled mark on screen. A sampled count is never presented as a total.
Some work is limited by design rather than sampled: within a trending topic, at most 200 media items an hour are fetched for examination, by default, taken in order of reach.
How much is sampled is itself reported. The share of trends that were sampled and the share of verdicts that were inconclusive are shown on the Health screen and in every report’s operations section. They are reported, not targeted.
What "inconclusive" protects you from
Inconclusive
On the Home screen: Could not tell
It protects you from saying more than the evidence holds: from a public claim that something is fake when no validated check found it, and from a claim that an item is clean when no validated check looked.
An inconclusive verdict always says why: no check ran, none that ran is validated for this kind of media, the validated checks disagreed, or the evidence fell between the lines. It is never turned into a negative.
In the first weeks of an installation, verdicts on the kinds of media whose checks have not yet passed are inconclusive. That is the intended behaviour, not a defect.
Tell us it is wrong
01
An analyst overturns a verdict
In the Review screen. If the verdict had been released, a correction follows.
02
Your organisation disputes a released item
A released verdict, report or reply that you dispute opens a correction.
03
Mark a flag as useful or noise
Each rule’s precision is shown from that feedback. A rule marked noise three times proposes a change of threshold to the policy owner.
04
Read how a correction is made
Who can ask, what is recorded and who is told.
Ask us about any limit on this page, and how it would show up in your own installation.
Request a demo/methodology/limits · Limits and known failure modes