# How to counter the liar’s dividend: procedures for checking and verifying information with AI

Author: Fernando Nieto Lobato
Original publication: 2025-11-12
Spanish original: https://estrategiabyaleph.substack.com/p/estrategia-111-cuando-ver-para-creer
English URL: https://elcontemplador.github.io/estrategia-english/essays/111/
Status: Published translation

This is a translation of the original Spanish essay published on 12 November 2025. Its claims, examples and forecasts retain that historical context.

English publication: 2026-09-29

*Archive note: this article preserves its November 2025 context, including figures, product availability and model limitations. Quotations are translated from the Spanish source unless an original English formulation is already present in the source or the accompanying image.*

**Disinformation is one of the most complex and serious problems facing the world today.** AI has enormous potential to make its spread easier and more powerful, but it also has considerable capacity to become a weapon against it.

[The World Economic Forum in Davos had already ranked disinformation as the leading short-term global risk in 2024](https://www.newtral.es/desinformacion-amenaza-foro-economico-mundial-davos-informe-riesgos-globales/20240115/). This concern is not exaggerated: [according to UNESCO, 85% of citizens worldwide are concerned about the impact of online disinformation, and 87% believe it has already significantly affected political life in their country](https://www.ipsos.com/es-mx/elecciones-y-redes-sociales-la-batalla-contra-la-desinformacion-y-los-problemas-de-confianza).

We have been considering an article on these questions for some time. Almost every month, at least one issue of estrategIA carries news about deepfakes or political disinformation involving AI. [The AI journalism course we are delivering through Institución Educativa ALEPH for the Aragón Journalists’ Association](https://periodistasdearagon.org/2025/10/10/periodistas-de-aragon-lanza-un-curso-avanzado-en-periodismo-local-e-inteligencia-artificial/), together with all the research we are undertaking in this field, has given us new perspectives and tools. This seems an excellent moment to share some keys to fighting disinformation with AI's help.

First, we should acknowledge an alarming development: generative AI has brought about a collapse in the “cost of producing a lie”. What once required a special-effects studio, a skilled voice actor or an army of propagandists can now be done with a mobile app. [A 2024 NewsGuard report documented a 1,000% increase in a single year in “content farms” — fake-news sites using AI to flood the web](https://maldita.es/malditatecnologia/20241003/granjas-contenidos-webs-anuncios-desinformacion-ia/).

The impact is multimodal and real, as these prominent examples show:

- **Audio — voice cloning:** [we saw robocalls using an audio deepfake of President Biden urging people not to vote in the New Hampshire primary](https://apnews.com/article/new-hampshire-primary-biden-ai-deepfake-robocall-f3469ceb6dd613079092287994663db5?utm_source=chatgpt.com).
- **Video — visual deepfakes:** [we witnessed a $25 million fraud at Arup, where an employee was deceived during a live video conference by deepfakes of the company's own executives](https://www.theguardian.com/technology/article/2024/may/17/uk-engineering-arup-deepfake-scam-hong-kong-ai-video?utm_source=chatgpt.com).
- **Text — hallucinations:** we still see large language models such as ChatGPT inventing “facts” and attributing them to reliable sources, a phenomenon the *New York Times* has bluntly called “disinformation”. Although newer AI models have substantially reduced hallucinations, this reminds us that human verification remains necessary, at least for now.

**Yet we believe the deeper danger is not the deepfake that fools us. It is the “[liar’s dividend](https://www.californialawreview.org/print/deep-fakes-a-looming-challenge-for-privacy-democracy-and-national-security)”.**

Legal scholars Robert (Bobby) Chesney and Danielle Keats Citron coined the term [in their 2019 *California Law Review* essay, “Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security”](https://www.californialawreview.org/print/deep-fakes-a-looming-challenge-for-privacy-democracy-and-national-security). They explicitly introduce “liar’s dividend” to describe a strategic gain for liars: as public awareness of synthetic fabrications grows, they can dismiss material that is actually genuine as “fake”.

As the public becomes aware that any video or audio recording might be false, a malicious actor — an unscrupulous politician, a campaign manager determined to win at any cost — gains the ability to dismiss *authentic* evidence, such as a genuine leaked recording or a real compromising video, as “just a deepfake”.

When audiovisual evidence loses its power, the fundamental principle of “seeing is believing” is eroded. If everything can be false, nothing is true.

## Basic strategies to help citizens avoid false stories

Before tackling more complex matters, I would like to share some basic practical advice for all of us as citizens. First, **be sceptical of the spectacular and the urgent**: false stories appeal to our emotions so that we share without thinking. Second, **look for the story in trusted media**. If no serious outlet reports the claim, it is suspicious. Third, make use of online resources and **consult established fact-checkers**: [Maldita](https://maldita.es/), [Newtral](https://www.newtral.es/fact-check/), [EFE Verifica](https://verifica.efe.com), [AFP Factual](https://factual.afp.com) and [VerificaRTVE](https://www.rtve.es/noticias/verificartve) in Spanish; [Snopes](https://www.snopes.com), [PolitiFact](https://www.politifact.com) and others in the English-speaking world. Alternatively, use a tool such as those discussed towards the end of this article, or AI as a verifier.

Fourth, **check the original source**. Who is making the claim? On which page or profile did it first appear? Examine the URL: many disinformation sites imitate real media outlets using almost identical domain names. Look for the site's “About” or “Who we are” section to assess its credibility. Fifth, **check images and videos**: accept nothing as genuine until you have subjected it to a reverse search or detailed inspection. Sixth, **do not rely on AI alone**. If you simply ask ChatGPT, “Is this true?”, it may answer yes or no and give reasons. But you must always corroborate its answer and question it more thoroughly, as we will see below.

[In a test conducted by RTVE more than two years ago, an early version of ChatGPT was effective at exposing false stories about COVID, but its knowledge had a time limit](https://www.rtve.es/noticias/20230127/prueba-chat-gpt-frente-desinformacion/2418106.shtml#:~:text=abiertas%20%28OSINT%29,respuestas%20y%20argumentos%20no%20difieren). AI models are not trained on up-to-the-minute information: their training normally has a cut-off several months before release, and they do not update themselves — at least for now, although papers describe promising techniques in this area. They therefore need to search the internet; they cannot verify current news using prior knowledge alone.

[![Six steps for detecting false stories and disinformation.](https://elcontemplador.github.io/estrategia-english/assets/images/0934ce75-6763-488d-8522-b16ba700ff44_1180x995.png)](https://substackcdn.com/image/fetch/$s_!Dhe_!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0934ce75-6763-488d-8522-b16ba700ff44_1180x995.png)

*English translation of the infographic, “Don't be taken in! Six key steps to detect false stories and disinformation”:*

1. **Question it.** Be suspicious of the spectacular and the urgent. False stories appeal to your emotions so that you share without thinking.
2. **Cross-check.** Look for the story in trusted media. If no serious outlet reports it, it is highly suspicious.
3. **Verify.** Consult fact-checkers such as Maldita, Newtral and EFE Verifica, or Snopes and PolitiFact in English.
4. **Investigate.** Examine the original source: who is making the claim? What does the URL look like? Find the site's “Who we are” section.
5. **Analyse.** Check images and videos. Run a reverse search to see whether they are old or taken out of context.
6. **Use AI — cautiously.** Do not trust it completely. AI can make mistakes or lack current data, because its knowledge has a cut-off date.

## Fighting fire with fire: using AI as a shield against disinformation

How do we fight this? The paradoxical answer is: with more AI.

Communicators and journalists will not be able to confront a deluge of machine-generated disinformation using manual tools. Traditional fact-checking, which takes hours or days, cannot compete with falsehoods generated in minutes.

Journalism must adopt an assisted-investigation model, using AI as a force multiplier. The key is to retain the human-in-the-loop model. AI does not replace the journalist; it strengthens their capabilities.

An interesting example of this partnership is [the newsroom at Newtral in Spain, which uses an internal AI system called Claim Hunter to monitor hundreds of politicians’ social-media accounts](https://www.newtral.es/periodismo-inteligencia-artificial-avances-newtral/20220624/). The AI does not decide what is true or false. It acts as a high-speed filter, detecting verifiable claims — data, figures and promises — and sending them to a Slack channel. The human team reviews those alerts, applies editorial judgement and decides what deserves investigation.

## Defensive procedures using generative AI

For journalists and communicators on the front line, generative AI — ChatGPT, Gemini, Claude and others — can be the most powerful research assistant ever created, or the most dangerous source of hallucinations. It all depends on how we use it.

Below are **five essential defensive procedures for integrating AI into your verification workflow, simply by using these prompts in your preferred AI tool**. We do, however, **strongly recommend advanced reasoning models with access to web search**. Otherwise, the results will probably be wrong.

### 1. A procedure for checking claims

Do not simply ask AI, “Is this true?” It will give you a generally unreliable summary. Demand evidence, with a prompt along these lines:

> Act as a professional fact-checker. Analyse this claim: “[CLAIM]”. Search the web for three (3) primary, highly authoritative sources — international media, academic studies or government reports — that corroborate or refute it. Provide a table with three columns: 1) the evidence provided, as a direct quotation; 2) the source, with its URL; and 3) your verdict: Verified, False or Misleading.

### 2. A procedure for comparing sources

Use AI's large context window to make exhaustive textual comparisons, a task at which AI already outperforms humans. Models such as Gemini 2.5 Pro, [available free through Google AI Studio](https://aistudio.google.com/prompts/new_chat), are suitable; otherwise, use the companies’ most advanced models. Free models almost always lack sufficient capacity to analyse lengthy documents.

> I will give you two articles — Text 1 and Text 2 — about the same event. Act as a research analyst. Your objective is to find discrepancies. Identify and list: 1) every factual point — names, dates and figures — on which they explicitly contradict each other; and 2) key information mentioned in one text but omitted from the other.

### 3. A procedure for detecting bias and framing

AI is excellent at sentiment analysis and at recognising patterns of language that may reveal a hidden agenda.

> Analyse the following article [TEXT]. Act as a media analyst specialising in bias. Identify: 1) loaded language — emotional or derogatory; 2) framing biases — what is emphasised and what is downplayed; and 3) exclusion biases — voices, perspectives or information relevant to the story that are conspicuously absent.

### 4. Checking a source's authority

It is essential to try to establish whether information comes from reliable sources. A prompt like this can help:

> Verify the existence and credentials of [PERSON/INSTITUTION MENTIONED]. Is this a real, trustworthy source? Give me a brief account of their background in relation to the issue under discussion. What authority do they have on this subject?

### 5. The devil's-advocate procedure

Use AI not only to check others, but to test the strength of your own investigation before publication.

> Example prompt: I am writing an article with the following thesis: “[THESIS]”. My three main pieces of evidence are [E1, E2, E3]. Act as a sceptical critic and a rigorous editor. What are the weakest points in my argument? What counterarguments could an opponent put forward? What additional evidence do I need to make my case irrefutable?

Alongside these five broadly “defensive” prompts against disinformation, this week's practical “Prompts for GPT-5” section in the original newsletter offers some more “offensive” ones that can be used directly to search for reliable information.

## The Swiss Army knife of information verification

Alongside large language models, which can be enormously helpful, we need specialist forensic tools to combat disinformation. Some highly advanced tools, such as Intel FakeCatcher or Resemble AI Detect, are paid products and usually available only to media organisations. Today, however, we bring you a genuine Swiss Army knife of verification that is easy to use and has surprised us with its versatility and simplicity. If you would like to explore further, [consult this excellent collection from Spanish public television's verification service](https://www.rtve.es/noticias/verificartve/herramientas-de-verificacion), [including an additional tab with more advanced tools](https://www.rtve.es/noticias/verificartve/herramientas-de-verificacion/avanzadas/).

[![English presentation of the InVID-WeVerify verification plugin.](https://elcontemplador.github.io/estrategia-english/assets/images/4e559440-acbf-4774-8d1b-0d07c18453fb_1600x685.png)](https://substackcdn.com/image/fetch/$s_!0VGJ!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4e559440-acbf-4774-8d1b-0d07c18453fb_1600x685.png)

*Original English text in the illustration, “The Verification Plugin”:*

> The Verification Plugin, also known as InVID-WeVerify, was co-developed by AFP and is designed as a verification “Swiss Army knife”.
>
> It has many useful features to help fact checkers, journalists and researchers with online investigations.

- **[InVID Verification Plugin](https://www.invid-project.eu/tools-and-services/invid-verification-plugin/).** This extremely powerful tool installs in a browser as a simple plugin. Several European research consortia have supported its development, with participants including AFP. Its particular value lies in bringing multiple analytical functions together in one interface. It helps us establish a video's context — platform, prominent comments and possible location — split it into images and search online to see whether those images were already circulating, and use a magnifier to read signs or examine details. It also reveals hidden file information, or metadata, including location where available, and lets us search X/Twitter within specified dates and places.

I think the best way to discover its capabilities is through an example such as the following.

*Archive note: the preserved source introduces an example here, but the export contains no accompanying video or recoverable link.*

[To learn how to get the most from the tool, we recommend this InVID Verification Plugin course from AFP and Google News Initiative, aimed primarily at journalists and researchers](https://digitalcourses.afp.com/courses/verification-plugin).

Bear in mind, however, that **deepfake detectors**, including those built into the plugin and even more advanced alternatives, **are not silver bullets**. Their results vary with the data and conditions. They can fail to generalise and are vulnerable to adversarial attacks: manipulation tools are specifically trained to become better at evading detectors. In the coming years, we will surely see a vigorous race between AI disinformation generators and advanced detection systems, also based on algorithms.

## Conclusion: technology and judgement against falsehoods

**Disinformation and deepfakes pose a serious problem for politics, government and journalism, but we are not defenceless. Just as malicious actors adopt AI to deceive, communicators and citizens can use it to expose them. The key is training and collaboration between people and technology.** An algorithm can scan thousands of posts in a minute, but a critical human eye is still needed to interpret results and make ethical decisions. In the fight against lies, AI supplies the muscle and speed; the journalist or communicator supplies contextual intelligence and responsibility.

From a broader strategic perspective, political leaders should advocate a clear, coherent regulatory framework that holds malicious actors accountable without suppressing freedom of expression. The response to disinformation cannot be unilateral. It requires coordinated action by government, platforms and civil society to create an environment in which truth has a chance to prevail.

In the age of generative AI, disinformation demands profound adaptation. For journalists, that means returning to the foundations of investigative journalism, equipped with new tools and a renewed commitment to verification. For communicators, it means more agile crisis management and a more deliberate effort to build trust. For politicians, it demands a more vigorous defence of truth and stronger institutional resilience. Success in this new environment will depend on everyone's ability to adapt, collaborate and, above all, maintain an unwavering commitment to truth and responsibility.

[Fernando Nieto Lobato](https://es.linkedin.com/in/fernandonietolobato?utm_source=chatgpt.com)

*Director of Digital Innovation at Institución Educativa ALEPH and Director of estrategIA*