AI is a system that repeats language patterns from training data, rather than necessarily reflecting truth or factual accuracy.
Hallucinate: 2023 word of the Year
Dictionary.com chose ‘hallucinate’ as its Word of the Year for 2023, as lookups and usage of this word increased significantly when referring to AI systems producing false or fabricated information.
But let’s dig a little deeper.
The term “hallucinate,” which is frequently used to describe AI’s incorrect outputs, is a misnomer that risks humanising these systems.
To humanise AI is to forget our responsibilities
It’s more than a semantic decision; it shapes our understanding of AI’s capabilities and limitations.
When we attribute human-like errors to AI, we unintentionally overestimate its perceived intelligence while underestimating the intricate, human-engineered processes that underpin it.
This could lead to a failure to recognise AI developers’ responsibilities.
Real-world consequences
Understanding AI’s limitations and capabilities is critical in my field, where the transition from dual to solo to uncrewed cockpits is underway.
We must keep in mind that at its core, AI is a tool that humans have created and improved.
It reflects the data it receives, so its ‘thought process’ is a reflection of our own inputs rather than an independent entity capable of human-like reasoning or errors.
This distinction is critical in aerospace, where incorporating AI into cockpit design demands precision, clarity, and a thorough understanding of AI’s role.
As we grow to trust AI-assisted cockpits, we must constantly refine our AI terminology and perceptions to match their actual function and capability.
This will both inform our approach to pilot training and evaluation in these AI-enhanced environments, and guide better engineering and design practices.
By using the correct terminology, data engineers can more easily pinpoint the design feature and correct it, leading to fewer iterations.
Potato-potahto
The following words are commonly used to describe situations where AI systems provide information that is incorrect, potentially due to flaws in their algorithms or training data.
I would like to see them more widely used in feedback mechanisms, deployment and operational testing.
For example by providing a pop-up list in the Chat-GPT window when you hover over the thumbs-down icon quickly allowing you to pinpoint the type of error.
Each term below (and by no means a complete list) highlights some of the more common potential issues within AI systems, underscoring the importance of comprehensive training, robust algorithm design, and ongoing evaluation to ensure AI models function as intended.
- Misleading: In AI, the term “misleading” refers to the output or conclusions generated by an AI system that lead to incorrect or deceptive understandings. This typically occurs when the AI’s responses or predictions give a false impression, either due to flawed processing of its input data or limitations in its algorithmic design. A misleading AI system might provide information that, while technically correct within its dataset, conveys an incorrect interpretation in a real-world context.
- Inaccurate: “Inaccurate” in AI context means the AI system’s outputs are incorrect or not exact. This can be due to errors in the data the AI was trained on, issues in the model’s design, or limitations in the AI’s ability to interpret complex or nuanced data. Inaccuracy implies a deviation from truth or correctness, which can range from minor errors to significant falsehoods in the information the AI provides.
- Biased: AI systems are described as “biased” when they display prejudice for or against one thing, person, or group compared to another, in a way considered to be unfair. Bias in AI often stems from the data it was trained on. If the training data contains biases—either due to the nature of the data collection process or inherent societal biases—the AI system can perpetuate and even amplify these biases in its outputs. This is a significant issue in AI ethics, as biased AI can lead to unfair or discriminatory outcomes.
- Unreliable: In AI, “unreliable” refers to systems that are not consistently accurate or dependable in their performance. An unreliable AI might provide correct outputs under certain conditions but fail under others. This inconsistency can be due to various factors, including the complexity of the task, limitations in the AI’s design, or variability in the input data. Unreliable AI systems are unpredictable in their accuracy, making them unsuitable for critical applications where consistent performance is essential.
- Erroneous: This term is used when AI outputs are wrong or incorrect. Erroneous outputs can be due to various factors, including algorithmic errors, misinterpretation of data, or limitations in the AI model’s processing capabilities. Imagine an AI system designed to predict aircraft maintenance needs. If it incorrectly identifies a fully functional engine as needing repair due to a misinterpretation of sensor data, this prediction would be erroneous.
- Overfitting: This is a term used to describe a situation where an AI model is too closely tailored to the specifics of its training data and fails to generalise well to new, unseen data. Overfitting leads to AI outputs that perform well on training data but poorly on real-world data (there are key exceptions with ACAS-Xu being a classic example). An AI model developed to simulate flight conditions might perform exceptionally well in scenarios it was trained on but fail to accurately simulate unexpected weather conditions, indicating it’s overfitted to its training data.
- Underfitting: The opposite of overfitting, underfitting occurs when an AI model is too simplistic and fails to capture the complexity of the training data. This leads to AI outputs that are generalised and lack accuracy. A traffic management AI that is too simplistic might fail to accurately predict traffic patterns on complex urban air mobility networks, showing signs of underfitting.
- Noisy: In AI, noisy outputs refer to results that are cluttered with irrelevant or extraneous information (manifesting in the same way as an under/overfitting model). This can happen when the AI model processes data that contains a lot of sensor noise, measurement errors, missing or duplicated data and outliers—random or irrelevant information that confuses the model.
- Ambiguous: AI outputs are termed ambiguous when they can be interpreted in multiple ways or lack clarity. This can be due to vagueness in the data the AI model has processed or limitations in the model’s ability to discern clear patterns.
- Conflicted: Sometimes, AI systems may provide outputs that seem to contradict each other or existing knowledge. This can happen due to conflicting data within the training set or the model’s inability to resolve discrepancies in the data. If an AI model for weather prediction provides one forecast indicating clear skies and another indicating storm conditions for the same time and location due to conflicting data sources, the outputs would be conflicted.
- Incomplete: This term is used when AI outputs are missing key information or insights. This can occur when the AI model has insufficient data or lacks the capability to fully process the complexity of the data.
Understanding these terms and their implications is crucial, especially in fields like aerospace where the precision and reliability of AI outputs are critical.
With its current trajectory, AI is set to transform our industry. To fully realise its potential, however, we must ground our understanding in reality rather than resorting to unrepresentative human-like tropes.
This is the key to realising AI’s true potential for improving human-machine collaboration, particularly in highly complex environments like ours.
Thanks for reading!
Please feel free to give it a clap if you enjoyed it. Alternately, offer it a small encouraging applause.
