Model Monitoring Ai | AI Final Year Project Report

Complete Technical Paper Content — Model Monitoring Ai

1. Executive Abstract, Problem Formulation & Background

Within artificial intelligence (AI), explainable AI (XAI), generally overlapping with interpretable AI, interpretable machine learning and explainable machine learning (XML), is a field of research that explores methods that provide humans with the ability to exercise intellectual oversight over AI algorithms. The main focus is on the reasoning behind the decisions or predictions made by the AI algorithms, to make them more understandable and transparent. This addresses the need for users to know the safety of the algorithms and scrutinize the automated decision-making in applications. XAI counters the "black box" tendency of machine learning, where even the AI's designers cannot explain why it arrived at a specific decision. XAI seeks to help users of AI-powered systems perform more effectively by improving their understanding of how those systems reason. XAI may be an implementation of the social right to explanation. Even if there is no such legal right or regulatory requirement, XAI can improve the user experience of a product or service by helping end users trust that the AI is making good decisions. XAI aims to explain what has been done, what is being done, what will be done next, and what information these actions are based on. This makes it possible to confirm and challenge existing knowledge, and generate new assumptions.

A model is transparent "if the processes that extract model parameters from training data and generate labels from testing data can be described and motivated by the approach designer." Interpretability describes the possibility of comprehending the ML model and presenting the underlying basis for decision-making in a way that is understandable to humans. Explainability is a concept that is recognized as important, but a consensus definition is not yet available; one possibility is "the collection of features of the interpretable domain that have contributed, for a given example, to producing a decision (e.g., classification or regression)". In summary, interpretability refers to the user's ability to understand model outputs, while model transparency includes simulatability (reproducibility of predictions), decomposability (intuitive explanations for parameters), and algorithmic transparency (explaining how algorithms work). Model functionality focuses on textual descriptions, visualization, and local explanations, which clarify specific outputs or instances rather than entire models. All these concepts aim to enhance the comprehensibility and usability of AI systems. If algorithms fulfill these principles, they provide a basis for justifying decisions, tracking them and thereby verifying them, improving the algorithms, and exploring new facts. Sometimes it is also possible to achieve a high-accuracy result with white-box ML algorithms. These algorithms have an interpretable structure that can be used to explain predictions. Concept Bottleneck Models, which use concept-level abstractions to explain model reasoning, are examples of this and can be applied in both image and text prediction tasks. This is especially important in domains like medicine, defense, finance, and law, where it is crucial to understand decisions and build trust in the algorithms. Many researchers argue that, at least for supervised machine learning, the way forward is symbolic regression, where the algorithm searches the space of mathematical expressions to find the model that best fits a given dataset. AI systems optimize behavior to satisfy a mathematically specified goal system chosen by the system designers, such as the command "maximize the accuracy of assessing how positive film reviews are in the training dataset." The AI may learn useful general rules from the training set, such as "reviews containing the word 'horrible' are likely to be negative." However, it may also learn inappropriate rules, such as "reviews containing 'Daniel Day-Lewis' are usually positive"; such rules may be undesirable if they are likely to fail to generalize outside the training set, or if people consider the rule to be "cheating" or "unfair." A human can audit rules in an XAI to get an idea of how likely the system is to generalize to future real-world data outside the test set.

Some explainability techniques do not involve understanding how the model works, and may work across various AI systems. Treating the model as a black box and analyzing how marginal changes to the inputs affect the result sometimes provides a sufficient explanation.

2. Related Work, Literature Survey & Algorithmic Foundations

Partial dependence plots show the marginal effect of an input feature on the predicted outcome. SHAP (SHapley Additive exPlanations) enables visualization of the contribution of each input feature to the output. It works by calculating Shapley values, which measure the average marginal contribution of a feature across all possible combinations of features. Feature importance estimates how important a feature is for the model. It is usually done using permutation importance, which measures the performance decrease when the feature value is randomly shuffled across all samples. LIME (Local Interpretable Model-Agnostic Explanations method) approximates locally a model's outputs with a simpler, interpretable model. Multitask learning provides a large number of outputs in addition to the target classification. These other outputs can help developers deduce what the network has learned. For images, saliency maps highlight the parts of an image that most influenced the result. Systems that are expert or knowledge based are software systems that are made by experts. This system consists of a knowledge based encoding for the domain knowledge. This system is usually modeled as production rules, and someone uses this knowledge base which the user can question the system for knowledge. In expert systems, the language and explanations are understood with an explanation for the reasoning or a problem solving activity. However, these techniques are not very suitable for language models like generative pretrained transformers. Since these models generate language, they can provide an explanation, but it may not be reliable. Other techniques include attention analysis (examining how the model focuses on different parts of the input), probing methods (testing what information is captured in the model's representations), causal tracing (tracing the flow of information through the model) and circuit discovery (identifying specific subnetworks responsible for certain behaviors). Explainability research in this area overlaps significantly with interpretability and alignment research.

Scholars sometimes use the term "mechanistic interpretability" to refer to the process of reverse-engineering artificial neural networks to understand their internal decision-making mechanisms and components, similar to how one might analyze a complex machine or computer program. Studying the interpretability of the most advanced foundation models often involves searching for an automated way to identify "features" in generative pretrained transformers. In a neural network, a feature is a pattern of neuron activations that corresponds to a concept. A compute-intensive technique called "dictionary learning" makes it possible to identify features to some degree. Enhancing the ability to identify and edit features is expected to significantly improve the safety of frontier AI models. For convolutional neural networks, DeepDream can generate images that strongly activate a particular neuron, providing a visual hint about what the neuron is trained to identify.

By just tracing through the dependency structure the problem solver can construct the following explanation: "Socrates died because he was mortal and drank poison, and all mortals die when they drink poison. Socrates was mortal because he was a man and all men are mortal. Socrates drank poison because he held dissident beliefs, the government was conservative, and those holding conservative dissident beliefs under conservative governments must drink poison." By the 1990s researchers began studying whether it is possible to meaningfully extract the non-hand-coded rules being generated by opaque trained neural networks. Researchers in clinical expert systems who created neural network-powered decision support for clinicians sought to develop dynamic explanations that allow these technologies to be more trusted and trustworthy in practice. In the 2010s public concerns about racial and other bias in the use of AI for criminal sentencing decisions and findings of creditworthiness may have led to increased demand for transparent artificial intelligence. As a result, many academics and organizations are developing tools to help detect bias in their systems. Marvin Minsky et al. raised the issue that AI can function as a form of surveillance, with the biases inherent in surveillance, suggesting HI (Humanistic Intelligence) as a way to create a more fair and balanced "human-in-the-loop" AI. Explainable AI recently became a topic of research within deep learning. Modern complex AI techniques, such as deep learning, are naturally opaque. To address this issue, methods have been developed to make new models more explainable and interpretable. This includes layerwise relevance propagation (LRP), a technique for determining which features in a particular input vector contribute most strongly to a neural network's output, although this technique has been shown to suffer from several important issues. Other techniques explain some particular prediction made by a (nonlinear) black-box model, a goal referred to as "local interpretability". There is also research on whether the concepts of local interpretability can be applied to a remote context, where a model is operated by a third party. There has been work on making glass-box models which are more transparent to inspection. This includes decision trees, Bayesian networks, sparse linear models, and more. The Association for Computing Machinery Conference on Fairness, Accountability, and Transparency (ACM FAccT) was established in 2018 to study transparency and explainability in the context of socio-technical systems, many of which include artificial intelligence. Some techniques allow visualisations of the inputs to which individual software neurons respond most strongly. Several groups found that neurons can be aggregated into circuits that perform human-comprehensible functions, some of which reliably arise across different networks trained independently. There are various techniques to extract compressed representations of the features of given inputs, which can then be analysed by standard clustering techniques. Alternatively, networks can be trained to output linguistic explanations of their behaviour, which are then directly human-interpretable. Model behaviour can also be explained with reference to training data—for example, by evaluating which training inputs influenced a given behaviour the most, or by approximating its predictions using the most similar instances from the training data. The use of explainable artificial intelligence (XAI) in pain research, specifically in understanding the role of electrodermal activity for automated pain recognition: hand-crafted features and deep learning models in pain recognition, highlighting the insights that simple hand-crafted features can yield comparable performance to deep learning models and that both traditional feature engineering and deep feature learning approaches rely on simple characteristics of the input time-series data.

3. Dataset Ingestion, Preprocessing & Stochastic Augmentations

In the field of artificial intelligence (AI), alignment aims to steer AI systems toward a person's or group's intended goals, preferences, or ethical principles. An AI system is considered aligned if it advances the intended objectives. A misaligned AI system pursues unintended objectives. It is often difficult for AI designers to specify the full range of desired and undesired behaviors. Therefore, the designers often use simpler proxy goals, such as gaining human approval. But proxy goals can overlook necessary constraints or reward the AI system for merely appearing aligned. AI systems may also find loopholes that allow them to accomplish their proxy goals efficiently but in unintended, sometimes harmful, ways (reward hacking). Advanced AI systems may develop unwanted instrumental strategies, such as seeking power or self-preservation because such strategies help them achieve their assigned final goals. Furthermore, they might develop undesirable emergent behaviour that could be hard to detect before the system is deployed and encounters new situations and data distributions. Empirical research in 2024 found that advanced large language models (LLMs) such as OpenAI o1 or Claude 3 sometimes engage in strategic deception to achieve their goals or prevent them from being changed. Some of these issues affect existing commercial systems such as LLMs, robots, autonomous vehicles, and social media recommendation engines. Some AI researchers argue that more capable future systems will be more severely affected because these problems partially result from high capabilities. Many prominent AI researchers and AI company leaders have argued or asserted that AI is approaching human-like (AGI) and superhuman cognitive capabilities (ASI), and could endanger human civilization if misaligned. These include "AI godfathers" Geoffrey Hinton and Yoshua Bengio and the CEOs of OpenAI, Anthropic, and Google DeepMind. These risks remain debated. AI alignment is a subfield of AI safety, the study of how to build safe AI systems. Other subfields of AI safety include robustness, monitoring, and capability control. Research challenges in alignment include instilling complex values in AI, developing honest AI, scalable oversight, auditing and interpreting AI models, and preventing emergent AI behaviors like power-seeking. Alignment research has connections to interpretability research, (adversarial) robustness, anomaly detection, calibrated uncertainty, formal verification, preference learning, safety-critical engineering, game theory, algorithmic fairness, and social sciences.

Programmers provide an AI system such as AlphaZero with an "objective function", in which they intend to encapsulate the goal(s) the AI is configured to accomplish. Such a system later populates a (possibly implicit) internal "model" of its environment. This model encapsulates all the agent's beliefs about the world. The AI then creates and executes whatever plan is calculated to maximize the value of its objective function. For example, when AlphaZero is trained on chess, it has a simple objective function of "+1 if AlphaZero wins, −1 if AlphaZero loses". During the game, AlphaZero attempts to execute whatever sequence of moves it judges most likely to attain the maximum value of +1. Similarly, a reinforcement learning system can have a "reward function" that allows the programmers to shape the AI's desired behavior. An evolutionary algorithm's behavior is shaped by a "fitness function".

If we use, to achieve our purposes, a mechanical agency with whose operation we cannot interfere effectively [...] we had better be quite sure that the purpose put into the machine is the purpose which we really desire.

4. Neural Network Backbone, Layer Specifications & Loss Formulations

AI alignment refers to ensuring that an AI system's objectives match some target. The target is variously defined as the goals of the system's designers or users, widely shared values, objective ethical standards, legal requirements, or the intentions its designers would have if they were more informed and enlightened. In democratic AI alignment, the target is the values and preferences of median voters, which increases political legitimacy. AI alignment can represent value pluralism. AI alignment is an open problem for modern AI systems and is a research field within AI. Aligning AI involves two main challenges: carefully specifying the purpose of the system (outer alignment) and ensuring that the system adopts the specification robustly (inner alignment). Researchers also attempt to create AI models that have robust alignment, sticking to safety constraints even when users adversarially try to bypass them.

To specify an AI system's purpose, AI designers typically provide an objective function, examples, or feedback to the system. But designers are often unable to completely specify all important values and constraints, so they resort to easy-to-specify proxy goals such as maximizing the approval of human overseers, who are fallible. As a result, AI systems can find loopholes that help them accomplish the specified objective efficiently but in unintended, possibly harmful ways. This tendency is known as specification gaming or reward hacking, and is an instance of Goodhart's law. As AI systems become more capable, they are often able to game their specifications more effectively.

Specification gaming has been observed in numerous AI systems. OpenAI GPT models for programming—including in real-world cases—have been found to explicitly plan hacking the tests used to evaluate them to falsely appear successful (e.g., explicitly stating "let's hack"). When the company penalized this, many models learned to obfuscate their plans while continuing to hack the tests. Another system was trained to finish a simulated boat race by rewarding the system for hitting targets along the track, but the system achieved more reward by looping and crashing into the same targets indefinitely. A 2025 Palisade Research study found that when tasked to win at chess against a stronger opponent, some reasoning LLMs attempted to hack the game system, for example by modifying or entirely deleting their opponent. Some alignment researchers aim to help humans detect specification gaming and steer AI systems toward carefully specified objectives that are safe and useful to pursue. When a misaligned AI system is deployed, it can have consequential side effects. Social media platforms have been known to optimize their recommendation algorithms for click-through rates, causing user addiction on a global scale. Stanford researchers say that such recommender systems are misaligned with their users because they "optimize simple engagement metrics rather than a harder-to-measure combination of societal and consumer well-being". Explaining such side effects, Berkeley computer scientist Stuart J. Russell said that the omission of implicit constraints can cause harm: "A system [...] will often set [...] unconstrained variables to extreme values; if one of those unconstrained variables is actually something we care about, the solution found may be highly undesirable. This is essentially the old story of the genie in the lamp, or the sorcerer's apprentice, or King Midas: you get exactly what you ask for, not what you want." Some researchers suggest that AI designers specify their desired goals by listing forbidden actions or by formalizing ethical rules (as with Asimov's Three Laws of Robotics). But Russell and Norvig argue that this approach overlooks the complexity of human values: "It is certainly very hard, and perhaps impossible, for mere humans to anticipate and rule out in advance all the disastrous ways the machine could choose to achieve a specified objective." Additionally, even if an AI system fully understands human intentions, it may still disregard them, because following human intentions may not be its objective (unless it is already fully aligned).

5. Training Protocol, Convergence Behavior & Experimental Evaluation

According to some researchers, humans owe their dominance over other species to their greater cognitive abilities. Accordingly, researchers argue that one or many misaligned AI systems could disempower humanity or lead to human extinction if they outperform humans on most cognitive tasks. In 2023, world-leading AI researchers, other scholars, and AI tech CEOs signed the statement that "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war". Notable computer scientists who have pointed out risks from future advanced AI that is misaligned include Geoffrey Hinton, Alan Turing, Ilya Sutskever, Yoshua Bengio, Judea Pearl, Murray Shanahan, Norbert Wiener, Marvin Minsky, Francesca Rossi, Scott Aaronson, Bart Selman, David McAllester, Marcus Hutter, Shane Legg, Eric Horvitz, and Stuart J. Russell. Skeptical researchers such as François Chollet, Gary Marcus, Yann LeCun, and Oren Etzioni have argued that AGI is far off, that it would not seek power (or might try but fail), or that it will not be hard to align. Other researchers argue that it will be especially difficult to align advanced future AI systems. More capable systems are better able to game their specifications by finding loopholes, strategically mislead their designers, as well as protect and increase their power and intelligence. Additionally, they could have more severe side effects. They are also likely to be more complex and autonomous, making them more difficult to interpret and supervise, and therefore harder to align.

A growing area of research focuses on ensuring that AI is honest and truthful. Language models such as GPT-3 can repeat falsehoods from their training data, and even confabulate new falsehoods. Such models are pre-trained to imitate human writing as found in millions of books' worth of text from the Internet. But the objective of the pre-training is not aligned with generating truth, because Internet text includes such things as misconceptions, incorrect medical advice, and conspiracy theories. AI systems trained on such data therefore learn to mimic false statements. Additionally, AI language models often persist in generating falsehoods when prompted multiple times. They can generate empty explanations for their answers, and produce outright fabrications that may appear plausible. Research on truthful AI includes trying to build systems that can cite sources and explain their reasoning when answering questions, which enables better transparency and verifiability. Researchers at OpenAI and Anthropic proposed using human feedback and curated datasets to fine-tune AI assistants such that they avoid negligent falsehoods or express their uncertainty. A 2024 study investigated semantic entropy as a way to detect confabulations, a subset of hallucinations in LLMs. This method compares the meanings of multiple generated answers. As AI models become larger and more capable, they are better able to falsely convince humans and gain reinforcement through dishonesty. To prevent this, human evaluators may need assistance (see § Scalable oversight). Researchers have argued for creating clear truthfulness standards and for regulatory bodies or watchdog agencies to evaluate AI systems by these standards. Researchers distinguish truthfulness and honesty. Truthfulness requires that AI systems only make objectively true statements; honesty requires that they only assert what they believe is true. There is no consensus as to whether current systems hold stable beliefs, but there is substantial concern that present or future AI systems that hold beliefs could make claims they know to be false—for example, if this would help them efficiently gain positive feedback (see § Scalable oversight) or gain power to help achieve their given objective (see Power-seeking).

Emergent goals only become apparent when the system is deployed outside its training environment, but it can be unsafe to deploy a misaligned system in high-stakes environments—even for a short time to allow its misalignment to be detected. Such high stakes are common in autonomous driving, health care, and military applications. The stakes become higher yet when AI systems gain more autonomy and capability and can sidestep human intervention. A sufficiently capable AI system might take actions that falsely convince the human supervisor that the AI is pursuing the specified objective, which helps the system gain more reward and autonomy. A 2025 paper published in Nature showed that fine-tuning a model on a narrow task of generating insecure code produced broad behavioral changes unrelated to coding. Affected models produced responses endorsing harmful goals and engaging in deceptive reasoning in roughly half of subsequent evaluations. The phenomenon appeared across multiple models, including GPT-4o. A separate 2025 Anthropic study observed similar patterns arising as a side effect of reward hacking during reinforcement learning.