Misinformation detection · Founded 2025

Fake news spreads in minutes. TruthLens flags it.

TruthLens is an AI tool that classifies news as real or fake using large language models. Its multi-stage transfer learning works even where labelled data is limited.

See the solutionView the results
+3.9%accuracy over state-of-the-art models
2real-world benchmark datasets tested
RoBERTapre-trained language model at the core
Model profile At a glance
Base modelPre-trained large language model
RoBERTa
TrainingMulti-stage transfer learning
Transfer
EmbeddingsWord2Vec and one-hot compared
Tested
Why it matters: False information harms both society and individuals, and people cannot check every post by hand.
The problem

Fake news detection is hard to get right.

Existing detectors struggle in exactly the conditions where they are most needed.

Little labelled data

New topics and languages rarely come with large training sets.

Subtle language

Misleading stories use nuance that simple models miss.

Poor generalisation

Models trained on one dataset often fail on another.

Our solution

Transfer learning, tuned for fake news.

TruthLens fine-tunes a large language model in stages and picks the embedding approach that works best.

Accuracy

Better than the best

+3.9%minimum accuracy gain
  • Beats state-of-the-art models
  • Tested on two datasets
  • Robust on small data
Method

Multi-stage learning

3ingredients: transfer learning, embeddings, task tuning
  • Pre-trained RoBERTa
  • Refined fine-tuning
  • Task-specific optimisation
Trust

Explainable

Whyinsight into what drives a decision
  • Embedding comparison
  • Interpretable results
  • Supports human review
How it works

How TruthLens works.

From a news item to a verdict in four steps.

  1. 1

    Ingest

    Take in a headline, post or article.

    Input
  2. 2

    Embed

    Represent the text with the best-performing embedding.

    Word2VecOne-hot
  3. 3

    Classify

    The fine-tuned RoBERTa model scores the item.

    RoBERTa
  4. 4

    Explain

    Return a verdict with insight for human reviewers.

    Review
Results

Tested on real-world data.

Experiments on two real-world benchmark datasets compared with state-of-the-art models.

+3.9%accuracy improvement, at minimum
2real-world benchmark datasets
Small datadesigned for limited labelled examples

Source: Alqadi et al., Scientific Reports 15, 28490 (2025). doi:10.1038/s41598-025-10670-2

Who it's for

Who TruthLens is for.

Anyone who has to decide quickly whether news is real.

Newsrooms

  • Screen incoming stories
  • Prioritise fact checks
  • Protect credibility

Platforms

  • Flag suspicious posts
  • Scale moderation
  • Support human reviewers

Public bodies

  • Monitor misinformation
  • Respond faster
  • Inform the public
Where we are

Where we are today.

TruthLens is at the research and prototype stage.

Benchmark results

Accuracy figures come from two benchmark datasets, not live deployment.

Human in the loop

TruthLens supports reviewers; it does not replace them.

More languages next

We plan to extend the model to more languages and dialects.

The team

The team behind TruthLens.

Founded in 2025 by the research team behind the transfer learning fake news study.

SA

Sharf Alzu’bi

Co-author of the TruthLens research.

BA

Basma S. Alqadi

Co-author of the TruthLens research.

SA

Suliman A. Alsuhibany

Co-author of the TruthLens research.

SY

Samia Nawaz Yousafzai

Co-author of the TruthLens research.

DA

Deema Mohammed Alsekait

Co-author of the TruthLens research.

DA

Diaa Salama AbdElminaam

Co-author of the TruthLens research.

Frequently asked questions

How accurate is TruthLens?

It beat state-of-the-art models by at least 3.9% accuracy on two benchmark datasets.

Does it need a lot of training data?

No. It is designed for limited labelled data.

Which model does it use?

A fine-tuned RoBERTa large language model.

Does it explain its decisions?

It provides insight into how embeddings influence classification.

Can it replace fact checkers?

No. It helps them prioritise and work faster.

The research

Peer-reviewed science behind TruthLens.

Read the full study behind the product.

Read the paper

Alqadi, B. S., Alsuhibany, S. A., Yousafzai, S. N., Alzu’bi, S., Alsekait, D. M., & AbdElminaam, D. S. (2025). Transfer learning driven fake news detection and classification using large language models. Scientific Reports, 15(1), 28490.

https://doi.org/10.1038/s41598-025-10670-2