The future of AI fact checking is in transparency, not grades.
The “Truth Score” Premise:* Why can’t we just put a 0–100 score on articles or live speeches like a paper being graded?
Let’s go ahead and dive into that premise to see why our goal should be tweaked from grades to transparency.
What is AI Fact Checking
From a technical standpoint, an AI system can analyze media and assign a confidence or credibility rating on a 0–100 scale. However, the AI does not test “truth” directly—it measures verifiability, source consistency, and digital authenticity.
- Articles & Written Media: Natural Language Processing (NLP) models extract individual claims (e.g., “Unemployment dropped by 3% in 2023”) and cross-references them against structured, verified knowledge bases like government statistics, court documents, and peer-reviewed journals. The score reflects how closely the claims match empirical data.
- Images & Photography: The system scans for digital manipulation through synthetic noise patterns, lighting inconsistencies, and edge artifacts common in Generative AI. It also checks cryptographic metadata (like C2PA provenance standards) and performs reverse-image searches to detect out-of-context reuse.
- Video & Deepfakes: Audio-visual AI checks frame-by-frame for facial synthesis anomalies, lip sync misalignment, and voice cloning signatures, while simultaneously evaluating the spoken transcript
for factual accuracy. - Advertisements: Ads combine factual claims with persuasive rhetoric. The AI evaluates explicit claims (e.g., “9 out of 10 doctors recommend”) against regulatory disclosures and clinical data, flagging misleading omissions or exaggerated framing.
Scoring Fact Checking Against Media Types
Real-time scoring is technically possible and already exists in early forms. Systems like Full Fact AI, Factiverse, and Duke University’s Squash project operate on live video streams using a three-step pipeline:
- Speech-to-Text Transcription: Live audio is transcribed in under two seconds.
- Claim Extraction: Natural language models instantly filter out opinions, greetings, and hyperbole to identify checkable factual statements.
- Database Matching: The claims are queried against verified fact-checks and official databases, generating a real-time confidence rating or lower-third screen overlay within 3 to 10 seconds of the candidate speaking. In a broadcast environment, this could appear as a dynamic meter or score next to a speaker’s name, accompanied by a brief citation.
Can Factual Accuracy be Hacked
Simple answer is YES. This is a major vulnerability known as “Data Poisoning” or a “Sybil Attack” on information networks. If an AI fact-checking system relies on general web indexing or search engine results to verify facts, an
attacker could deploy thousands of automated AI bots to publish articles, blog posts, and forum entries repeating the same false statistic. However, robust verification systems are built specifically to resist this:
- Source Weighting & Authority Allowable Listing: Advanced truth-scoring systems do not count web pages like a popularity vote. They assign high weight to cryptographically verified, primary sources—such as direct government API feeds (e.g., Bureau of Labor Statistics), judicial rulings, peer-reviewed scientific journals, and established wire services. Ten million fake blog posts would carry near-zero weight against a single primary database record.
- Knowledge Graph Consistency: AI engines rely on structured knowledge graphs (networks of interconnected real-world entities and dates). Injecting a fake fact usually creates mathematical contradictions within the broader graph, causing the system to flag the source network as compromised.
- The Risk: If an attacker succeeds in corrupting or compromising a primary database itself (or if the AI relies on search-engine consensus rather than primary sources), the truth score can be manipulated.
Real World Examples
Let’s take a look at some real world examples for using AI to verify facts.
Factiverse Live Fact-Checking
Used by Tjekdet (Denmark), Viestimedia (Finland), The Kyiv Independent. Factiverse streams live audio from political broadcasts, turns it into text in real time, isolates claim-heavy sentences, and cross-references them against multi-engine databases and verified fact-checking archives. During major political broadcasts (including EU Parliament debates and US
Presidential debates), newsrooms used Factiverse to surface citations and claim verdicts within minutes, giving journalists speed and coverage they could not match manually.
Full Fact AI
Over 40 fact-checking organizations across 30 countries. Full Fact AI monitors live TV broadcasts, speeches, and social feeds using BERT
based models fine-tuned to categorize claims (e.g., economic statistics vs. predictions vs. cause-and effect claims). During national elections, the system runs continuously in newsrooms. When a politician repeats a previously debunked claim on live television or during a speech, the AI instantly alerts newsrooms that the speaker has repeated a known falsehood, complete with the corresponding archive proof.
Checkmate
Used by the BBC, *Deutsche Presse-Agentur (dpa), The Sun, and The Times. A collaborative project built specifically for newsroom production teams. It
processes live stream video by chunking audio, transcribing it in near-real-time using speech models, and using LLMs to extract testable claims. Checkmate queries the claims against the Google Fact Check Explorer API and alerts studio journalists via a dashboard with context cards and source links, allowing producers to feed corrected facts directly to on-air anchors.
Duke Reporters Lab (Squash)
Duke reporters. others know as Squash is used in research newsrooms and by debate monitoring teams. Squash* ingests live closed-caption feeds or speech-to-text audio streams from televised debates. The AI compares spoken sentences against thousands of previously published fact checks from organizations like FactCheck.org, PolitiFact, and the Washington Post. When a candidate speaks a recognized claim during a debate, Squash pops up the matching fact-check report on a side-screen monitor within seconds.
Future Direction
So, is there a future for using AI as a fact checker? Yes, but not in the way most people originally think.
If the technology is used as a singular ‘truth score’ assigned to news, it will likely fail to build public trust. However, if implemented as a ‘interactive, transparent verification layer’, it has the potential to fundamentally transform how we consume media.
AI fact checking won’t be some meter sitting on top of your screen or news article, instead it will evolve into distinct architectures.
- Back of the house AI assistants (Newsroom co-pilots) – This is happening today as news agencies use AI to transcribe live debates query government databases in seconds, and surface previous quotes to anchors via earpieces or monitors. This makes journalists sharper and faster without putting an automated algorithm in charge of editorial decisions.
- Personal fact-checking extension layer – Instead of news networks scoring themselves, the future belongs to user-controlled AI tools. Imagine a browser extension or TV app that doesn’t tell you what to think, but lets you highlight any sentence spoken on air and instantly displays:
- The official source data (e.g., direct links to government, academic, or court records).
- Context that was omitted.
- How opposing viewpoints interpret the exact same statistic.
Below will outline why a ‘truth score’ isn’t the magic bullet for trust and what could actually restore public confidence.
Why This Won’t Restore Public Trust
It is tempting to think a 0–100 score would solve fake news, but media studies and psychological research show that a single numeric score often backfires:
1. The “Whose AI Is It?” Problem
The core issue with public trust today isn’t just a lack of facts; it is a deep skepticism of institutions.
- If a left-leaning network displays a high truth score for a politician, right-leaning viewers will claim the AI was programmed with partisan bias.
- If a right-leaning network displays a low truth score for an opposing policy, left-leaning viewers will call the system compromised.
- Result: Rather than trusting the score, the public simply debates the credibility of the AI, deepening polarization.
2. The Illusion of Objectivity on Subjective Issues
A score works well for simple math (e.g., “Did GDP grow by 2%?”). But politics and news are rarely that simple.
- If a politician says, “Our economy is failing working-class families,” that isn’t a strict true/false claim—it is a value judgment based on specific economic indicators. Assigning a score like “42/100” to an opinion or prediction projects a false sense of scientific certainty onto subjective debates.
3. “Implied Truth” Exploitation
When AI systems grade articles, bad actors quickly figure out how to game the algorithm. A bad faith writer can construct an article where 90% of the sentences are technically true statistics, but the overall conclusion is completely misleading. The AI might award it an “A” grade based on factual
coverage, giving a dishonest narrative an official seal of approval.
What Will Restore Public Trust
If a 0–100 score won’t fix trust, how can AI actually help? Research shows public trust returns when AI shifts from being an Auditor (telling you what is true) to an Enabler of Transparency (showing you the evidence).
- Exposing Provenance, Not Judgments: Instead of declaring “This image is fake,” AI tools are beginning to display cryptographic metadata (like C2PA standards) showing when the camera took the photo, if photoshop was used, and where it original appeared. People trust verifiable receipts more than black-box grades.
- Instant Context Engines: Trust is restored when AI acts like an instant index. When a speaker cites a cherry-picked stat, the AI doesn’t call them a liar; it simply displays the full 10-year trend line next to them on screen. It empowers the viewer to spot the spin themselves.
- Cross-Spectrum Comparison: Instead of one score, an AI layer could show: “Here is how 5 major news outlets across the political spectrum are reporting this exact same event.” Showing the full picture builds credibility far faster than claiming to possess the absolute truth.
Future Timeline
Roadmap to migrate from fragmented tools to a future implementations.
Near Term (Present – 2 Years)
Newsrooms integrate and standards emerge. AI live transcription and context dashboards become standard in political control rooms.
Hardware integration becomes common in consumer phones and broadcast cameras.
Medium Term (2-5 Years)
Focus is strictly to assist human journalists. Consumer facing context layers with browser and smart TV extensions allow users to click a claim and see primary source data in real time.
AI Moves past simple true/false to flag omissions, cherry picking (choosing only partial facts in an attempt to prove a point) and out of context statistics.
Long Term (5+ Years)
Open source, decentralized verification models emerge to combat ‘Whose AI’ bias. This results in interoperable truth ecosystem. Real time augmented reality/heads up display overlays during live broadcasts. This will all see the rollout of cryptographic verification standards across almost all digital media.
Our end result is the ‘truth score’ concept is fully abandoned in favor of interactive, instant source tracing.
Issues Ahead
All technology has it’s issues and AI is no exceptions. AI is smart and so are people. As the technology rolls out, unfortunately people being people will get better at lying in attempts to side step the AI transparency. Here’s four examples.
Weaponized Factuality Will Replace Outright Lies
As the models make simple factual lies instantly catchable, political rhetoric will evolve rather than become more honest. We will enter an era of weaponized factuality. Instead of making things up, public figures will rely entirely on true statements arranged in deeply dishonest ways. If every single sentence in a speech is 100% technically accurate according to government statistics, an automated fact-checker passes it with flying colors—even if the overall narrative is a complete distortion. The battleground will shift entirely from truth vs. falsehood to framing vs. context.
Moving from Authority to Traceability
For the past century, public trust relied on authority: “You should believe this because the New York Times, the President, or an expert institution said so.” That model is largely broken. Real-time AI verification represents a fundamental shift toward traceability: “Don’t take my word for it—click here to inspect the raw database record, watch the unedited video snippet, or review the original source code yourself.” The institutions that survive and thrive in the AI era won’t be the ones demanding trust based on their brand, but the ones making their receipts easiest to inspect.
Real-Time AI Will Flatten Human Speech
If politicians and public figures know that an AI co-pilot is listening to every word and instantly feeding counter-stats to an interviewer, speech will become far more sanitized, defensive, and scripted. While this eliminates blatant fabrications, it also risks making public discourse feel clinical and robotic. Spontaneity, figurative language, and rhetorical passion often rely on hyperbole. When hyperbole gets continuously flagged as “statistically imprecise” in real time, speakers will default to safe, pre-vetted corporate speak.
Bottleneck Isn’t Technology, It’s Human Psychology
We often talk about real-time fact-checking as a technical problem waiting for a software solution. But the deepest hurdle is cognitive psychology: most people do not consume political media to be informed; they consume it to feel validated. When presented with real-time facts that contradict their deeply held identity or worldview, humans rarely say, “Ah, thank you AI, I stand corrected.” Instead, we experience cognitive dissonance and actively seek reasons to dismiss the source. Until we account for how human beings process uncomfortable truths, even the most perfect, unbiased verification system in the world will only ever be a partial solution.
Examples of Issues
Let’s take a closer look at how we could sidestep our AI to make our point using factual data, just manipulated a little bit. Remember, all true statements, just rearranged or incomplete.
Example 1: Public Health Crisis
Ok, let’s set the stage. Imagine a viral outbreak occurs. Someone wanting to manipulate this for their benefit (say the vaccine provided is killing people) is aware that AI will check the facts. Here are the known facts.
- Over 10,000 people who received treatment X died this year. That is true in a country with 300 million. 10,000 people who happened to take the drug died of various natural causes, old age or advanced illness.
- The health ministry issued an urgent emergency alert regarding treatment X this morning. True, the ministry issued a routine warning about a single contaminated batch in one city.
- Emergency rooms in major hospitals are currently operating at maximum capacity. Again, true, but it’s also the winter flu season which always pushes up local ER’s to capacity
Now our bad actor walks up to the microphone and delivers this report:
Emergency rooms in major hospitals are currently operating at maximum capacity [Fact C]. This comes as the Health Ministry issued an urgent emergency alert regarding Treatment X this morning [Fact B], after records confirmed that over 10,000 people who received Treatment X died this year
[Fact A].
See the problem? All the ‘facts’ are accurate, just moved around to deliver a point the bad actor wanted to make. The AI ‘sees’ the facts and flags it as true, while the population (psychology manipulation) interprets this differently. The population thinks Emergency rooms are overflowing because 10,000 people died from treatment X. It’s so bad the government issued a warning about it.
Example 2: Percentages vs Raw Numbers
Claim: Violent crime is up 100% this year in city X.
Reality: Incidents went from 1 incident last year to 2 incidents this year in a city of 500,000 people. The “100%” figure is mathematically true, but using a percentage instead of raw numbers inflates a negligible change into a panic.
Reads likes someone or some group is selling fear.
Example 3: Cherry Picked Topic
Claim: Under my opponent’s policy, gas prices jumped 40%!
Reality: Gas prices rose 40% over a specific 3-month window due to a seasonal refinery pause, but dropped 15% overall across the full 4-year term. Both numbers are true; selecting the 3 month window weaponizes the data.
Root Causes of Issues
The bad news is AI is having trouble catching that. The good news is that it is being worked on. Let’s take a look at 4 root causes of the issues.
- Comparative Context (Baselines and Denomination)
- Claim: City had 500 violent crimes this year!
- Missing Context: City A has 2 million residents (low rate). City B had 400 violent crimes, but has only 50,000 residents (high rate).
- AI’s Job: Automatically append the per-capita baseline.
- Casual Context (Correlation vs Causation)
- Claim: Ice cream sales rose and drowning rates surged at the exact same time.
- Missing Context: Both are caused by a third variable, summer heat.
- AI’s Job: Detect when two true, concurrent facts are being spuriously linked to imply cause and effect.
- Proportional Context (Scale):
- Claim: Government department spent $50 million on redundant software.
- Missing Context: $50 million represents .001% of that departments annual operating budget.
- AI’s Job: Show the figure relative to the total budget so the audience understands its true scale.
- Sarcasm:
- A speaker rolls his eyes and says ‘Oh, brilliant plan’.
- Issue: The AI is literal and records ‘Oh, brilliant plan’ as the speaker praising the plan.
- The AI needs to learn sarcasm through a multi-modal shift.
- Acoustic Pitch – Sarcasm in spoken audio usually involves prolonged vowels, a lower pitch, or exaggerated cadence.
- Visual Cues – Computer vision scans for micro-expressions like raised eyebrows, asymmetrical smirks, or eye rolls that contradict the spoken words.
- Semantic Incongruity – The AI compares the statement against the immediate situation. If a politician stands in front of a flooded highway and says, “Great weather today,” the AI flags the clash between the visual background and the literal words.
Summary
We shouldn’t expect artificial intelligence to act as a supreme judge that decides what we ought to believe. The true value of AI fact-checking isn’t that it hands out grades—it’s that it gives the audience a magnifying glass. Ironing out context and sarcasm will take time, but giving everyday viewers instant access to the underlying facts is a necessary, game-changing start.
In closing I would also say, do your homework. If there’s something you’re passionate about, don’t take the word of your politicians, local vendors or even family and friends. Put in the leg work to determine if you’re given half-truths, cherry picked stats or outright falsehoods. It’s a shame it’s come to this, I get that. Keep the faith though. If the media and our politicians won’t give us the truth, perhaps AI can push them into that direction.