Feature
AI is becoming increasingly prone to confidently delivering incorrect information, or hallucinations. A recent NewsGuard report revealed the rate of false claims generated by top AI chatbots nearly doubled within a year, climbing from 18% in August 2024 to 35% in August 2025 when responding to news-related prompts. This rise coincides with models being programmed to provide more answers — thanks to internet access — rather than declining to respond when uncertain.
To understand what’s driving this spike and how it might be curbed, we spoke with AI engineers, product leaders, data strategists and academic researchers about the underlying mechanisms, risks and emerging strategies to tackle this challenge.
Table of Contents
- The Misinformation Spike
- How Are AI Hallucinations Measured?
- The Chatbots That Lie the Most
- What’s Fueling the Hallucination Surge?
- Experts Weigh In on the Spike
The Misinformation Spike
The scale of AI-generated misinformation is surging at an unprecedented pace, according to NewsGuard’s August 2025 report — an acceleration that’s drawing concern from both the public and policymakers.
AI Hallucination Rates: August 2024 vs August 2025
| Year | Hallucination Rate | Change |
|---|---|---|
| 2024 | 18% | N/A |
| 2025 | 35% | +94% |
Why does this matter? The real-world stakes are enormous. As AI-generated content floods social media, news sites and messaging platforms, the risks multiply: voters are exposed to fabricated stories ahead of elections, patients encounter bogus health advice (in fact, more than 1 in 5 Americans follow AI medical advice that's later proven wrong) and trust in both AI technology and the media ecosystem continues to erode. In short, the cost of unchecked hallucinations is no longer theoretical — it’s affecting the way people perceive reality and make critical decisions.
The dramatic rise in these errors isn’t just a technical glitch; it’s a challenge that sits at the intersection of AI design, data quality and the social responsibility of those using the technology.
Related Article: Reducing AI Hallucinations: A Look at Enterprise and Vendor Strategies
How Are AI Hallucinations Measured?
To understand the scale of AI hallucinations, NewsGuard conducted a systematic analysis across major generative AI platforms over the past year. Their methodology involved inputting hundreds of prompts into leading AI chatbots — including OpenAI’s ChatGPT, Google Gemini and Microsoft Copilot — and then rigorously fact-checking the outputs for accuracy.
The evaluation covered a number range of real-world scenarios, including:
- Current events
- Health guidance
- Historical facts
- Breaking news
Any output that included fabricated details, misleading claims or “invented” citations was classified as a hallucination. This approach provided a quantifiable measure of how frequently each model produced false or misleading content.
The Chatbots That Lie the Most
The findings showed a significant increase in hallucination rates across the board, but especially among models used in high-traffic applications. By comparing results year-over-year, the report found that the problem is not only persistent, but growing more severe as AI models become more deeply embedded in everyday information flows.
The AI chatbots with the highest hallucination rates include:
- Inflection (56.67%)
- Perplexity (46.67%)
- Meta (40%)
- ChatGPT (40%)
- Copilot (36.67%)
While all large language models are prone to hallucinations to some degree — an acknowledged limitation of current technology — the doubling in frequency is considered unusually high by industry standards. For context, previous benchmarks from academic studies often noted modest improvements with each generation of models. The current spike reverses that trend, raising new questions about how these systems are trained and governed.
What’s Fueling the Hallucination Surge?
Recent tests show hallucination rates are actually increasing in newer language models. According to OpenAI’s own technical report, its o3 model hallucinated 33% of the time and o4-mini 48%, compared to just 16% for its o1 model from late 2024.
Importantly, this trend isn’t limited to OpenAI. Public leaderboards such as Vectara’s have found that other prominent models, like DeepSeek-R1, also experienced double-digit increases in hallucination rates over their predecessors — even as they were engineered to show more reasoning steps.
Key Factors Driving AI Hallucinations
| Factor | Description | Impact |
|---|---|---|
| Wider Model Coverage | Models are designed to answer more questions | Increased risk of confident, inaccurate outputs |
| Reduced Refusals | AI is less likely to say "I don't know" | Fewer safe fails; more plausible fabrications |
| Training Data Quality | Models are exposed to more unvetted data sources | Hallucinations reflect real-world misinformation |
| Topic Sensitivity | Hallucinations are more common in fast-moving topics like politics or health | Higher risk in critical domains |
Several factors driving the recent spike in AI hallucinations include:
- Modern language models are designed to tackle a wider range of questions than before, often opting to answer even when the information is incomplete or unreliable.
- Many leading platforms have reduced guardrails in an effort to appear more helpful and conversational, but this increases the risk of generating convincing but inaccurate responses.
- Models are now frequently trained on vast, real-time internet sources — many with questionable accuracy — resulting in higher exposure to misinformation.
Related Article: Navigating AI Hallucinations: A Personal Lesson in Digital Accuracy
Experts Weigh In on the Spike
"Nowadays, the bulk of models are trained to 'answer better', not 'answer truthfully.'"
- Rory Bokser
IoT Expert & Head of product, Moken
"The demand to 'always respond' means that systems will now produce something convincing, but inaccurate or unverifiable when evidence is thin. Mix in real-time news — where data is in rapid flux, and you raise the odds for error," observed Jonathan Garini, enterprise AI strategist and CEO at Fifth Element.
Rory Bokser, AI/IoT expert and head of product at Moken, echoed that sentiment, adding, "Nowadays, the bulk of models are trained to 'answer better', not 'answer truthfully.' In other words, you’re incentivizing models to make confident estimates when you fine-tune them for lower rejection rates, faster completions and smoother conversation flow."
This shift in priorities — toward always providing an answer, even in uncertain scenarios — has become an industry-wide trend.
Webinar
Sep
25
CMS Briefing: A Live Look at What’s Next in AI-Driven Platforms
Learn how leading organizations are using AI‑driven tools to publish faster, personalize smarter and stay secure.
Register
Webinar
Oct
15
Agentic AI Playbook Real-World Customer Service Use Cases You Can Deploy Now
Boost self-service by 30% and slash call volume by 63% with agentic AI.
Register
Webinar
On demand
Ready or Not: How Data-First Organizations Are Unlocking Agentforce Potential
Learn how to cut through the noise, activate Agentforce and build a Salesforce AI strategy that actually delivers.
Watch Now
Webinar
On demand
AI in Customer Service: Faster Resolutions, Happier Customers
Don’t let rising demand burn out your team. See how to build a smarter, more resilient support org.
Watch Now
Webinar
On demand
From Hype to High-Impact CX Strategies That Actually Scale
Turn buzzworthy AI and outsourcing trends into measurable CX wins with fresh 2025 data.
Watch Now
Webinar
On demand
Insights to Action Rethinking the Contact Center for Real Business Impact
Join our exclusive webinar to hear CX executives share their innovative strategies for transforming service delivery.
Watch Now
Model architecture also plays a role. Developers are training LLMs on ever-larger and more diverse datasets, much of it scraped from the open internet. This increases coverage and fluency, but also exposes models to outdated, low-quality or biased information.
"Rapid expansion of access to internet data + weak verification of sources — models are connected to more updated information, but not always from reliable or verified sources," said Taras Tymoshchuk, CEO at Geniusee.
All told, the surge in hallucinations is less a technical glitch and more a byproduct of how today’s AI models are designed, trained and deployed, with tradeoffs between breadth, helpfulness and factual reliability still unresolved.
FAQs
According to NewsGuard’s 2025 report, the AI platforms with the highest hallucination rates include:
- Inflection (56.67%)
- Perplexity (46.67%)
- Meta (40%)
- ChatGPT (40%)
- Copilot (36.67%)
They can spread misinformation about elections, health, law and current events, eroding trust in media and technology. For example, more than 1 in 5 Americans follow AI health advice later proven wrong.
Surprisingly, newer models often perform worse. OpenAI’s o3 model hallucinated 33% of the time, while o4-mini rose to 48% — compared to just 16% for its earlier o1 model.
Key drivers of AI hallucinations include:
- Reduced refusal rates (less “I don’t know”)
- Broader question coverage
- Training on unvetted internet sources
- Sensitivity to fast-changing topics like politics and health
Vendors and researchers are experimenting with approaches such as grounding responses in verified databases, improving model transparency, training on curated datasets and introducing stronger refusal mechanisms.
Signs of hallucination include:
- Overly confident answers without sources
- Fabricated citations
- Statistics that don’t appear in reputable outlets
Verifying claims through trusted sources remains essential.
