Algorithms vs Algorithms: AI in the Fight Against Advanced Digital Disinformation
To stop the falsehoods generated by Artificial Intelligence, another Artificial Intelligence is needed. In this "algorithmic arms race," those who produce fake
We have entered the era in which truth is a scalable concept. If Generative Artificial Intelligence has broken down the costs and time needed to create perfect fake news and deepfake videos, we can no longer defend ourselves using exclusively the manual work of journalists and fact-checkers. The time debt is insurmountable: a network of bots can publish one hundred thousand false news items in the time it takes a human editor to verify just one.
To survive this "information war," we are developing a digital immune ecosystem. Today, the same neural architectures used to generate falsehoods are being retrained to recognize them.
In this in-depth analysis, we will explore the new frontier of algorithms designed to counter disinformation. We will analyze the most recent frameworks adopted by the European Commission (including VERA.ai and AI4TRUST), deepfake identification technologies, and the ethical challenges that arise when we ask an algorithm to decide what is true and what is false.
1. The Detection Paradigm: NLP, Bots, and Anomalies
Modern disinformation is almost never artisanal; it is an industrial operation. Therefore, defense must be based on identifying industrial patterns.
A systematic analysis published on ScienceDirect on the use of AI based on trilateral frameworks against disinformation highlights that algorithms focus on three vectors: the Actor (who publishes), the Content (what is said), and the Platform (how it spreads). Natural Language Processing (NLP) systems do not merely search for suspicious keywords; they analyze the emotional tone, syntactic urgency, and lack of logical complexity typical of automatic generators.
The World Economic Forum (WEF) reiterates the importance of AI in combating online disinformation through Anomaly Detection. If a Twitter (X) account inactive for two years suddenly starts tweeting every three minutes in seven different languages on a geopolitical topic, AI blocks the account preemptively, recognizing it as a hostile bot even before its content is analyzed.
This acceleration of deception is a vital issue for our democracies. We explored the logic behind these campaigns in our special feature on Fake News and AI: The Era of Information Warfare.
2. The Arms Race Against Deepfakes
The hottest front is undoubtedly that of synthetic media. Falsified videos and audio have the power to manipulate financial markets or elections in real time.
Today, threat intelligence companies like Sensity AI are deploying their Deepfake Detection Hub. These systems do not watch the video with "human eyes"; they apply a multi-layer analysis. The algorithm does not look for macroscopic errors but analyzes the invisible: anomalies in pixel density, mathematically inconsistent lighting, or the absence of micro blood pulsations on the subject's skin.
Audio is an even more insidious attack vector (Voice Cloning). Cybersecurity companies like McAfee have launched tools such as the Deepfake Audio Detector, based on DNN transformers. These models operate in real time (for example, during a call or while watching an online video) to calculate the frequency and harmonics of the voice, alerting the user if the voice of their presumed "relative in distress" has been computer-synthesized.
3. The European Union's Response: VERA.ai and AI4TRUST
Faced with the scale of the phenomenon, the European Union has allocated massive funds through the Horizon program to develop open source tools available to journalists.
The journal Frontiers in Political Science published a mapping of AI use in counter-disinformation, examining dozens of European projects. Projects like VERA.ai and AI4TRUST stand out, whose goal is not to replace the human fact-checker (as happens on automated commercial platforms), but to provide them with a "cognitive exoskeleton." AI scours the multilingual web, verifies the origin of images (automated reverse image search), and cross-references politicians' statements with official parliamentary databases in milliseconds.
In Italy too, centers of research excellence like the Fondazione Bruno Kessler (FBK) actively participate. In a recent whitepaper on progress and challenges of generative AI against disinformation, FBK researchers highlight the importance of real-time multilingual monitoring. European hoaxes often originate in Russia, are tested in Eastern Europe, translated into English for global social media, and finally arrive in Italy. AI makes it possible to trace the "patient zero" of a fake news story before the local public opinion becomes infected.
4. Ethical Limits: Who Controls the "Truth Machine"?
There is a dangerous limit in entrusting the management of public "Truth" to opaque neural networks. If we allow Google, Meta, or X to use Machine Learning (ML) algorithms to moderate and censor content on a global scale (as analyzed by the portal Recentiprogressi in the essay Does AI amplify or combat disinformation?), we are creating a monopoly on reality.
The main risk is Over-blocking and Bias. An algorithm trained in America might not grasp Italian political irony or sarcasm, labeling a piece of satire as "disinformation" and removing it (a form of automated algorithmic censorship). Even worse, as highlighted by the portal IA2023 on algorithmic challenges to truth and the role of European responses (DSA), blindly relying on these automatic fact-checkers can turn into a weapon against political dissidents, depending on who holds the servers and sets the criteria for "falsity."
This risk is intrinsically linked to the nature of training data. We explored the genesis of these errors in our in-depth analysis on Algorithmic Bias, AI, and Invisible Discrimination.
Key Operational Takeaways
- Detecting Frameworks: Current defenses do not limit themselves to analyzing text. They evaluate the provenance (where it comes from), the amplification (bot networks retweeting simultaneously), and the behavioral anomaly of the profile to identify the inorganic nature of a message.
- Audio Deepfake at the Center: New commercial tools are shifting from video (already widely covered) to audio (Voice Cloning), implementing real-time alerts on smartphones to prevent "relative in danger" scams.
- The European "Human-in-the-Loop" Approach: Projects like VERA.ai and FBK reject total automatic censorship. They use AI to flag and cross-reference data on a large scale, leaving the final action of debunking or blocking to the journalist (or forensic analyst).
- The Censorship Bias Paradox: The rush to automate fact-checking to block rigged elections risks generating "false positives," blocking legitimate dissenting opinions or political satire. The fairness of the control algorithm remains the central challenge of 2026.
FAQ: Understanding Defense Against Algorithmic Disinformation
1. Why can't we simply use "human intelligence" (journalists) to debunk fake news? Because of the asymmetry of scale. A GenAI software can produce thousands of fake articles and spread them across thousands of fake blogs in an hour. To rigorously debunk one single complex piece of news, a journalist needs hours or days (Asymmetric Debunking). Without algorithms to quickly identify what to filter, the journalistic system collapses under the pressure of data.
2. What is "Anomaly Detection" applied to Social Media? It is a mathematical technique used by platforms to find Bot networks (Botnets). If AI detects that 5,000 accounts, all created last September, simultaneously share the same image with the same text at 3 AM, it identifies the action as a coordinated non-human operation (a statistical anomaly) and deactivates the network.
3. How can an algorithm tell that a video is a Deepfake if it looks real to me? Our brain is fooled by the coherence of the face and expression. The detection algorithm, on the other hand, splits the video frame by frame. It can analyze metadata (missing cryptographic signatures), frequency variations of light on pixel edges, or the absence of rPPG (remote photoplethysmography), i.e., the tiny, invisible pulsation of skin color due to the heartbeat, impossible for an AI generator to replicate.
4. What is the European Digital Services Act (DSA)? It is the European legislative package that legally obliges large platforms (Very Large Online Platforms like Facebook, TikTok, X) to take stringent technical measures to curb the spread of disinformation and deepfakes on their networks. If a platform does not mitigate these risks, Europe can fine it billions of dollars in percentages of its global turnover.
5. What is meant by "Assisted Fact-Checking"? In opposition to "fully autonomous" fact-checking (where AI deletes posts by itself), "assisted" fact-checking (developed by projects like AI4TRUST) uses AI as a co-pilot. The journalist queries the news item, and AI returns a dashboard in seconds with: the origin of the photo, historical precedents of the news in other countries, and references to official documents, exponentially speeding up human work.
Conclusions: An Endless Race
The fight against disinformation in the era of Artificial Intelligence is not a problem with a definitive solution. It is a "Cat and Mouse Game".
Every time companies like Sensity AI or consortia like VERA.ai train a new algorithm to detect a deepfake, the disinformation network uses that same algorithm to understand where their fake video was "discovered," correct the defective pixels, and generate a perfect version 2.0 that is untraceable by the old filter.
In this relentless race, technology is necessary to avoid drowning in synthetic chaos, but the utopia of delegating the search for "Truth" to machines proves dangerous. The ultimate defense against disinformation will never come from silicon, but from the cognitive resilience of human education and the critical thinking of citizens.
Bibliographic References and Sources
To ensure technological, academic, and institutional accuracy, this article drew upon the following primary sources:
- Strategic Frameworks and EU Projects:
- ScienceDirect – AI-based digital disinformation: A trilateral framework (Actor, Content, Platform). Link
- Frontiers in Political Science – The use of AI in counter-disinformation (Mapping of Horizon projects like VERA.ai and AI4TRUST). Link
- Fondazione Bruno Kessler (FBK) – Generative AI and disinformation: multilingual progress and challenges. Link
- Detection and Security Tools (Deepfake):
- Ethical Risks, Bias, and DSA: