Baptista is a multifaceted hybrid professional known for her work on the future of work. She is the CEO of FoReal HR Services. Building a team of an efficient & effective workforce is her business.
Affecting lives is her calling! She is an HR Generalist, International Development Expert, Public Speaker, Researcher, and Lifestyle Interventionist. You can reach her @Sarahtistagh across all platforms.
This week, WhatsApp’s AI assistant experienced a glitch, mistakenly sharing users’ numbers – Ops!. Despite this, Meta CEO Mark Zuckerberg praised AI, calling it ‘the most intelligent AI assistant you can freely use. What is your experience?
Then a recent MIT study explored how AI tools like ChatGPT affect students’ cognitive engagement. Researchers asked participants to write SAT-style essays using either ChatGPT, Google Search, or no assistance. Those who used Chat-GPT showed the lowest brain activity and creativity, often relying on copy-paste strategies and producing formulaic, “soulless” essays.
The study, though not yet peer-reviewed, raises concerns about AI’s impact on critical thinking and learning, especially for younger users, warning that early reliance on generative AI could hinder brain development, urging caution before integrating such tools into early education.
In recent years, artificial intelligence (AI) systems—especially large language models—have dazzled the world with their ability to generate text, engage in conversation, and synthesize complex topics. Yet as their presence grows, so too does awareness of one of their most critical limitations: hallucinations. Far from science fiction visions or psychedelic experiences, in the realm of AI, “hallucinations” refer to instances when an AI system confidently produces information that is factually incorrect, misleading, or entirely fabricated.
Understanding why these hallucinations happen, identifying examples where they’ve gone awry, and developing strategies to mitigate them is essential for the responsible adoption and integration of AI technologies, especially in sensitive sectors like human resources, healthcare, law, journalism, and education.
In humans, we refer to hallucinations as perceptions or experiences that occur without any external stimulus. They can involve any of the senses, including: Visual hallucinations – seeing things that aren’t there (e.g., shapes, objects, people). Auditory hallucinations – hearing sounds or voices that aren’t real. Tactile hallucinations – feeling sensations on the skin that aren’t actual. Olfactory hallucinations– smelling odors that aren’t present, and Gustatory hallucinations – tasting flavors that aren’t real.
Hallucinations in humans can be caused by various factors, such as Mental health conditions, as seen in schizophrenia, bipolar disorder, or severe depression. Neurological disorders as in Parkinson’s disease, Alzheimer’s disease, or epilepsy. Substance use with certain drugs or alcohol. Sleep deprivation due to lack of sleep or disrupted sleep patterns, and in medical conditions as seen in high fever or infection.
Can A. I Lie or Hallucinate too?
An AI hallucination is said to occur when a model generates output that doesn’t align with reality or factual data. Unlike a human lying deliberately, the model is said to produce what seems plausible based on patterns in its training data.
These hallucinations often arise because AI models do not “know” facts the way humans do. Rather, they predict what comes next in a sequence of words based on patterns and probabilities learned during training. This predictive approach is powerful for generating coherent and contextually appropriate text, but it also opens the door to confidently presented errors.
Notable Examples of AI Hallucinations
A recent incident involving Meta’s WhatsApp AI helper has sparked privacy concerns after it mistakenly shared a private phone number with a user seeking a train service helpline. Instead of providing the correct contact for TransPennine Express, the AI gave out the number of James Gray, a property executive. When challenged, the AI gave contradictory explanations—first claiming the number was fictional, then suggesting it was generated based on patterns, and later implying it may have come from a database.
The affected user, Barry Smethurst, described the experience as “terrifying,” raising alarms about AI reliability and data protection. Meta responded by stating the number was publicly available and similar to the intended helpline, while acknowledging that the AI may return inaccurate outputs. The case highlights broader concerns about AI-generated misinformation and the need for stronger safeguards. Think about this – if A.I today is able to generate your mobile number, what happens tomorrow, perhaps with your bank accounts – that could be generated too?
Legal Fabrications, Historical and Biographical Errors
One of the most infamous AI hallucination cases occurred in 2023, when a lawyer in the United States used ChatGPT to generate legal research for a court filing. The AI model cited six court cases—none of which existed. The judge, upon reviewing the submission, expressed disbelief at the fictitious citations, resulting in penalties for the lawyer and intense scrutiny over the use of generative AI in legal proceedings.
AI models have also been found to generate fabricated quotes, incorrect dates, or even entirely fictional events when asked about historical figures. For instance, some users experimenting with early versions of language models received responses claiming that famous figures had written books or made statements they never did. These hallucinations are especially dangerous in educational contexts, where misinformation can easily become embedded in learners’ understanding.
A Norwegian man filed a complaint after ChatGPT falsely and confidently told him he was a convicted murderer, claiming he had killed two of his children. The information was found to be entirely fabricated. What is your own experience so far- using A.I, and what have you found? From the Safe Space Community, some have realized that inaccurate or outdated A.I information, contextual misunderstandings, limited domain-specific knowledge, ambiguity, and unclear prompts could be some of the issues impacting response accuracy and relevance from these A..I’s
Scientific Missteps and Healthcare Inaccuracies
In academic and scientific settings, hallucinations can be problematic. Language models have been known to produce plausible-looking scientific abstracts, including citations to non-existent studies. This poses challenges for researchers, editors, and peer reviewers who might mistakenly assume such output to be legitimate.
Healthcare AI applications—such as those involved in diagnosis or patient communication—can be particularly risky when hallucinations occur. An AI suggesting an incorrect medication, providing an inaccurate prognosis, or hallucinating symptoms could have serious consequences. Even if designed for informational purposes only, these tools can shape decision-making in vulnerable situations. Be Agile!
Why Do Hallucinations Happen?
AI hallucinations are rooted in the fundamental mechanics of how language models are trained. Lack of ground truth, statistical prediction, training data noise, ambiguous prompts, and overconfidence by design, and many more could be contributing to these hallucinations.
AI models are trained on vast amounts of text, not on a live, up-to-date database of facts. They don’t cross-check with reality unless explicitly connected to a knowledge base or search engine. These models generate the most likely sequence of words, which often sounds correct, but might not be correct.
The training data itself may include false or misleading information, which the model learns as plausible. If a prompt is vague or asks for something speculative, the model may “fill in” with made-up details. Many language models are tuned to produce responses in a confident tone to sound more natural and engaging, even when the underlying information may be incorrect. Let’s watch these developments and be guided as such.
Strategies to Address AI Hallucinations
Hallucinations may never be eliminated based on what we are discovering so far as some causes, but their occurrence can be significantly reduced and better managed. Here’s how;
One of the most effective ways to minimize hallucinations is to ground AI outputs in real-time information. This can involve integrating search capabilities or connecting models to verified databases. For example, some AI systems can now cite sources directly from the web or structured datasets.
Reinforcement Learning from Human Feedback (RLHF) – AI developers increasingly use human feedback to train models not only to produce accurate information but also to recognize when to say, “I don’t know.” RLHF involves having human reviewers rate AI-generated responses, which helps fine-tune the models over time.
With Prompt Engineering – Users can improve the accuracy of AI output through careful prompt construction. By being explicit in their requests and specifying that factual sources be cited, users can guide the model toward more accurate answers. Example: Instead of asking, “What are the symptoms of a rare disease?” users might ask, “Summarize CDC-listed symptoms of [disease] based on up-to-date sources.”
Developers are working on ways for AI systems to express uncertainty more appropriately. Rather than declaring all answers with the same level of confidence, future models might indicate degrees of certainty, allowing users to weigh the information accordingly, promoting Confidence Calibration
Human-in-the-Loop Systems – In high-risk domains like medicine, finance, or law, it’s critical to have human experts oversee and verify AI-generated content. AI should augment—not replace—human judgment.
Opening AI models to external audits, evaluations, and red-teaming (purposefully trying to break the model) can reveal vulnerabilities and hallucination patterns. Transparency and auditing in model architecture and limitations help stakeholders use AI more responsibly.
Let’s encourage the use of citations and source attribution. When AI responses include citations or links to sources, it enhances accountability. Some platforms now include citation features that trace information back to reliable origins, helping users verify claims independently.
Beyond Technical Fixes: A Culture of Critical Thinking
Addressing hallucinations isn’t just about algorithmic improvement—it’s about fostering digital literacy and critical engagement. Users, especially in educational or decision-making roles, must be trained to question and verify AI outputs.
Institutions should incorporate guidelines for ethical AI use, much like citation rules or plagiarism policies. In workplaces, policies around “AI-assisted writing” should clarify the boundaries between human accountability and algorithmic support.
Conclusion: The Path Ahead
As I consider as a Future of Work Expert, together with other thought leaders on how AI intersects with global priorities like education reform, digital transformation, and responsible innovation, hallucinations serve as a vivid reminder that technology must be approached with nuance.
AI is a powerful co-pilot—but not an infallible oracle. Its limitations highlight the continued importance of human judgment, diverse perspectives, and robust systems of verification. By investing in smarter model design, improving human-computer collaboration, and cultivating critical thinking, we can navigate the illusions that AI sometimes casts—and harness its strengths more safely and responsibly.
