Home News What Are LLM Hallucinations? Causes, Ethical Concern, & Prevention

What Are LLM Hallucinations? Causes, Ethical Concern, & Prevention

What Are LLM Hallucinations? Causes, Ethical Concern, & Prevention

Large language models (LLMs) are artificial intelligence systems able to analyzing and generating human-like text. But they’ve an issue – LLMs hallucinate, i.e., make stuff up. LLM hallucinations have made researchers anxious in regards to the progress on this field because if researchers cannot control the end result of the models, then they can not construct critical systems to serve humanity. More on this later.

Generally, LLMs use vast amounts of coaching data and complicated learning algorithms to generate realistic outputs. In some cases, in-context learning is used to coach these models using only just a few examples. LLMs have gotten increasingly popular across various application areas starting from machine translation, sentiment evaluation, virtual AI assistance, image annotation, natural language processing, etc.

Despite the cutting-edge nature of LLMs, they’re still liable to biases, errors, and hallucinations. Yann LeCun, current Chief AI Scientist at Meta, recently mentioned the central flaw in LLMs that causes hallucinations: “Large language models do not know of the underlying reality that language describes. Those systems generate text that sounds nice, grammatically, and semantically, but they don’t really have some type of objective aside from just satisfying statistical consistency with the prompt”.

Hallucinations in LLMs

Image by Gerd Altmann from Pixabay

Hallucinations seek advice from the model generating outputs which are syntactically and semantically correct but are disconnected from reality, and based on false assumptions. Hallucination is one among the key ethical concerns of LLMs, and it could actually have harmful consequences as users without adequate domain knowledge begin to over-rely on these increasingly convincing language models.

A certain degree of hallucination is inevitable across all autoregressive LLMs. For instance, a model can attribute a counterfeit quote to a celeb that was never said. They might assert something about a selected topic that’s factually incorrect or cite non-existent sources in research papers, thus spreading misinformation.

Nonetheless, getting AI models to hallucinate doesn’t at all times have antagonistic effects. For instance, a brand new study suggests scientists are unearthing ‘novel proteins with a vast array of properties’ through hallucinating LLMs.

What Causes LLMs Hallucinations?

LLMs can hallucinate resulting from various aspects, starting from overfitting errors in encoding and decoding to training bias.


Image by janjf93 from Pixabay

Overfitting is a problem where an AI model suits the training data too well. Still, it cannot fully represent the entire range of inputs it might encounter, i.e., it fails to generalize its predictive power to latest, unseen data. Overfitting can result in the model producing hallucinated content.

Encoding and Decoding Errors

Image by geralt from Pixabay

If there are errors within the encoding and decoding of text and its subsequent representations, this can even cause the model to generate nonsensical and erroneous outputs.

Training Bias

Image by Quince Creative from Pixabay

One other factor is the presence of certain biases within the training data, which may cause the model to present results that represent those biases somewhat than the actual nature of the info. This is comparable to the dearth of diversity within the training data, which limits the model’s ability to generalize to latest data.

The complex structure of LLMs makes it quite difficult for AI researchers and practitioners to discover, interpret, and proper these underlying causes of hallucinations.

Ethical Concerns of LLM Hallucinations

LLMs can perpetuate and amplify harmful biases through hallucinations and may, in turn, negatively impact the users and have detrimental social consequences. A few of these most vital ethical concerns are listed below:

Discriminating and Toxic Content

Image by ar130405 from Pixabay

For the reason that LLM training data is usually stuffed with sociocultural stereotypes resulting from the inherent biases and lack of diversity. LLMs can, thus, produce and reinforce these harmful ideas against disadvantaged groups in society.

They’ll generate this discriminating and hateful content based on race, gender, religion, ethnicity, etc.

Privacy Issues

Image by JanBaby from Pixabay

LLMs are trained on an enormous training corpus which frequently includes the non-public information of people. There have been cases where such models have violated people’s privacy. They’ll leak specific information equivalent to social security numbers, home addresses, mobile phone numbers, and medical details.

Misinformation and Disinformation

Image by geralt from Pixabay

Language models can produce human-like content that seems accurate but is, in reality, false and never supported by empirical evidence. This might be accidental, resulting in misinformation, or it could actually have malicious intent behind it to knowingly spread disinformation. If this goes unchecked, it could actually create antagonistic social-cultural-economic-political trends.

Stopping LLM Hallucinations

Image by athree23 from Pixabay

Researchers and practitioners are taking various approaches to handle the issue of hallucinations in LLMs. These include improving the range of coaching data, eliminating inherent biases, using higher regularization techniques, and employing adversarial training and reinforcement learning, amongst others:

  • Developing higher regularization techniques is on the core of tackling hallucinations. They assist prevent overfitting and other problems that cause hallucinations.
  • Data augmentation can reduce the frequency of hallucinations, as evidenced by a research study. Data augmentation involves augmenting the training set by adding a random token anywhere within the sentence. It doubles the scale of the training set and causes a decrease within the frequency of hallucinations.
  • OpenAI and Google’s DeepMind developed a method called reinforcement learning with human feedback (RLHF) to tackle ChatGPT’s hallucination problem. It involves a human evaluator who continuously reviews the model’s responses and picks out probably the most appropriate for the user prompts. This feedback is then used to regulate the behavior of the model. Ilya Sutskever, OpenAI’s chief scientist, recently mentioned that this approach can potentially resolve hallucinations in ChatGPT: “I’m quite hopeful that by simply improving this subsequent reinforcement learning from the human feedback step, we will teach it to not hallucinate”.
  • Identifying hallucinated content to make use of for instance for future training can also be a way used to tackle hallucinations. A novel technique on this regard detects hallucinations on the token level and predicts whether each token within the output is hallucinated. It also includes a way for unsupervised learning of hallucination detectors.

Token-level Hallucination Detection

Put simply, LLM hallucinations are a growing concern. And despite the efforts, much work still must be done to handle the issue. The complexity of those models means it’s generally difficult to discover and rectify the inherent causes of hallucinations appropriately.

Nonetheless, with continued research and development, mitigating hallucinations in LLMs and reducing their ethical consequences is feasible.

If you desire to learn more about LLMs and the preventive techniques being developed to rectify LLMs hallucinations, try unite.ai to expand your knowledge.


Please enter your comment!
Please enter your name here