External Side-Effects Across Disciplines

• External side-effects are unintended impacts that extend beyond a system’s primary scope, observable in programming (I/O operations), language (anaphora), and biomedical (drug interactions) settings.
• They are formally modeled using techniques like monadic lifting and dynamic semantics to capture state changes, context updates, and network propagation.
• Practical assessments in AI safety, pharmacovigilance, and physical systems highlight trade-offs and the need for evaluation benchmarks that measure both primary benefits and collateral risks.

External side-effects are effects whose impact extends beyond the immediate scope of the entity being analyzed: beyond the lexical scope of an expression, beyond the scope of a program itself, beyond the target tissue or target concept, or beyond the intended therapeutic or regulatory action. In programming languages, they include actions such as state change, exceptions, and interaction with the outside world by “printing documents, making sounds, operating robotic limbs” (Marsik et al., 2015). In natural language semantics, analogous phenomena include discourse anaphora and presupposition projection, which similarly transcend lexical boundaries (Marsik et al., 2015). In pharmacology, external side effects can arise from drug–drug interactions, network-mediated perturbation spread, or post-market patient experiences not captured at approval time (Zitnik et al., 2018, Perez-Lopez et al., 2015, Bartal et al., 2024). In machine learning and AI safety, the term is used for unwanted consequences that are not encoded in the primary objective, including collateral damage to neighboring concepts, attribute leakage, or irreversible environmental change (Turner et al., 2020, Saha et al., 20 Aug 2025). Across these literatures, the common structure is a distinction between a primary value or target and additional effects that propagate through a larger context.

1. Conceptual range across disciplines

A recurring formulation treats side effects as actions that “go beyond the lexical scope of an expression” or “even beyond the scope of the program itself” (Marsik et al., 2015). In the programming-language setting, the relevant boundary is the local expression or the program’s internal computation. External effects therefore include I/O operations and interaction with the outside world. In natural language, the relevant boundary is the sentence or phrase, while pragmatic phenomena such as discourse anaphora and presupposition projection alter a broader discourse context (Marsik et al., 2015).

In biomedical work, the boundary is typically the intended target of a treatment. Polypharmacy side effects “emerge because of drug-drug interactions,” and are “distinct from the individual side effects of the drugs themselves” (Zitnik et al., 2018). Network pharmacology introduces an additional notion of externality: some side effects are not directly caused by on-target inhibition or activation, but rather by “network ripple effects” in the human interactome (Perez-Lopez et al., 2015). Social-media pharmacovigilance extends the scope further, treating patient-reported experiences as a source of adverse side effects that may appear “outside controlled clinical trial environments” and outside structured surveillance systems (Bartal et al., 2024).

In AI safety, side effects denote unintended environmental changes not captured by the specified reward. Rewarding an agent for making a widget may be easy, but penalizing the “multitude of possible negative side effects” is hard; this motivates methods that discourage irreversible changes relative to a no-op baseline (Turner et al., 2020). In generative modeling, concept erasure produces side effects on “neighboring concepts,” permits “evasion of targets,” and induces “attribute leakage” (Saha et al., 20 Aug 2025). In video editing, ROSE defines side effects as the object’s “observable influence” on the environment, including shadows, reflections, light, translucency, and mirror effects (Miao et al., 26 Aug 2025).

This breadth suggests that “external side-effects” is not a domain-specific label but a structural one: an effect is external when it cannot be adequately characterized within the local object, action, or representation alone.

2. Formal representations of effects beyond local scope

A central formal device is type lifting. In programming languages, a pure value may have type $A$, while a computation that modifies state has type

$\text{State} \to (A \times \text{State}).$

For general effects, Moggi’s monadic abstraction gives a computation of value type $A$ the type $M~A$ (Marsik et al., 2015). For external effects such as I/O, the relevant form is $\text{IO}~A$; in Haskell, IO Int is an integer produced as the result of an I/O computation (Marsik et al., 2015). The point is that the denotation no longer consists only of a value, but of a value together with effectful interaction.

Dynamic semantics uses the same shift. Instead of plain propositions, sentences denote functions from contexts to propositions and updated contexts: $$\text{Context} \to (\mathbf{Prop} \times \text{Context}),$$ or, in monad-like form, $M~\mathbf{Prop}$ (Marsik et al., 2015). The paper explicitly states that “the type of a proposition changes into a function from discourse contexts to propositions and updated discourse contexts in order to handle anaphora,” paralleling the shift from values to state-transforming computations in programming languages (Marsik et al., 2015). The associated split between value and effect mirrors Stalnaker’s content/context distinction: truth-conditional content remains the value, while context update or presupposition triggering is treated as the effect (Marsik et al., 2015).

Logic-based work on contextual abduction formalizes side-effects as secondary consequences of explanatory hypotheses. Under weak completion semantics, explanations can depend on prior abductive context through inspection points $inspect(A)$ and $inspect_{\neg}(A)$, which can be abduced only if the corresponding hypothesis has already been abduced elsewhere (Pereira et al., 2014). The paper introduces formal definitions of necessary contextual side-effects, possible contextual side-effects, contested contextual side-effects, contextual relevant consequences, and jointly supported contextual relevant consequences (Pereira et al., 2014). Here the externality is not state or I/O, but dependence on abductive context beyond the primary observation.

In complexity-controlled computation, side effects are represented through regions and depth stratification. The modal $\lambda$-calculus with multithreading and side effects uses regions $\text{State} \to (A \times \text{State}).$0, read and write operations, and parallel composition $\text{State} \to (A \times \text{State}).$1 (Madet, 2012). The operational rules $\text{IO}~A$1 model consuming reads and cumulative writes (Madet, 2012). A depth system assigns each region a depth and constrains reads and writes to that depth, while shallow-first reduction preserves the observable order of side effects under multithreaded call-by-value evaluation (Madet, 2012). This is a formal response to a familiar problem: once effects can propagate through shared regions, evaluation order becomes part of the semantics.

3. Pharmacological and biomedical external side-effects

Drug safety work uses “side effects” in the clinical sense, but several papers distinguish effects that arise outside the intended therapeutic mechanism. In polypharmacy, side effects emerge from drug–drug interactions. Decagon represents drugs and proteins as nodes, and drug–drug interactions are labeled by side effect type; there are 964 distinct side effects represented as edge types (Zitnik et al., 2018). The task is multirelational link prediction: $\text{State} \to (A \times \text{State}).$2 and Decagon reports performance gains “outperforming baselines by up to 69%” (Zitnik et al., 2018). The formulation makes an external side effect explicit as a relation that appears only at the level of a drug pair, not at the level of either drug alone.

Network analysis of the human interactome offers a mechanistic interpretation. “Drug targets are generally better spreaders of perturbations than non-target proteins,” and “targets of drugs with side effects are even better spreaders than those of drugs without reported side effects” (Perez-Lopez et al., 2015). Spreading efficiency is quantified by Silencing Time and Perturbation Reach, computed with the Turbine software on a human interactome containing 12,439 proteins and 174,666 edges (Perez-Lopez et al., 2015). The paper interprets higher spreading efficiency as a reason targeted proteins can induce secondary impacts on the broader cellular network, increasing the likelihood of undesired side effects, especially those arising via network-mediated propagation rather than direct target action (Perez-Lopez et al., 2015).

Post-market surveillance increasingly treats public online discourse as a source of externally reported adverse side effects. For GLP-1 receptor agonists, one study integrated $\text{State} \to (A \times \text{State}).$3, Reddit, PubMed, SIDER, manufacturer-reported side effects, and ChatGPT, using a pre-trained NER model, ScispaCy en_ner_bc5cdr_md, to extract biomedical entities (Bartal et al., 2024). It identified 134 adverse side effects across all datasets, with 21 side effects found exclusively via social media analytics, including irritability and numbness (Bartal et al., 2024). Effectiveness was evaluated with the Overlap score

$\text{State} \to (A \times \text{State}).$4

with a maximum Overlap of 0.53 (Bartal et al., 2024). A later Reddit study on semaglutide and tirzepatide analyzed 410,198 posts, found 67,008 users self-reporting use, and reported that 43.5% described at least one side effect; gastrointestinal symptoms predominated, while reproductive symptoms and temperature-related complaints emerged as “unrecognized potential effects” (Sehgal et al., 12 Mar 2026).

A related LLM-based study on semaglutide used Reddit to construct a medication–side-effect knowledge graph $\text{State} \to (A \times \text{State}).$5, with medication nodes, side-effect nodes, and directed edges $\text{State} \to (A \times \text{State}).$6 containing severity, duration, dosage, and description (Duan et al., 6 Apr 2025). The pipeline yielded 7,225 relations and consolidated 2,284 raw side effect entities down to 96 standardized entities, with side effect extraction accuracy of 79.34% and severity extraction accuracy of 86.85% on a manually evaluated 5% sample (Duan et al., 6 Apr 2025). Comparison with FAERS showed a Spearman correlation in rank-frequency of 0.423 for the top 20 Reddit side effects (Duan et al., 6 Apr 2025).

A different predictive line uses only molecular structure. GNN–MGSEP formulates drug side-effect prediction as a multi-class, multi-label graph-focused classification over molecular graphs, with a binary target vector of length 2055 (Bongini et al., 2022). The recurrent GNN is trained with binary cross-entropy and Adam, and reports best results of approximately 95.25% binary accuracy, 86.73% AUC, and 29.13% AUPR (Bongini et al., 2022). By contrast, NRBdMF treats therapeutic effects and side effects as bidirectional aspects of the same drug-effect relation, assigning side effects $\text{State} \to (A \times \text{State}).$7, indications $\text{State} \to (A \times \text{State}).$8, and unknown or contradictory relations $\text{State} \to (A \times \text{State}).$9 (Azuma et al., 2022). Its enrichment score separates side effects at the top and indications at the bottom of the ranking, reducing false positives such as ranking antihypertensive drugs as likely causes of hypertension (Azuma et al., 2022).

A common misconception is that external side effects in pharmacology are identical to off-target toxicity. The interactome and polypharmacy work indicates a broader picture: interaction structure, propagation dynamics, and post-market behavioral context can all generate clinically relevant effects that are not reducible to a single target-binding narrative (Zitnik et al., 2018, Perez-Lopez et al., 2015, Bartal et al., 2024).

4. External side-effects in AI systems and generative models

In reinforcement learning, the canonical problem is that the reward function does not encode the full set of harms. Attainable Utility Preservation modifies the task reward by penalizing shifts in the optimal value of auxiliary reward functions relative to the no-op action: $A$0 The objective is to preserve the agent’s ability to optimize a wide range of rewards and thereby discourage irreversible side effects (Turner et al., 2020). In SafeLife, a large environment based on Conway’s Game of Life, AUP with a single unsupervised auxiliary reward function reduced disturbance of protected “green” cells by over 50% relative to PPO alone while maintaining or exceeding primary task reward; in the append-spawn task, AUP achieved 111% of the reward and only 39% of the side effects compared to PPO after 5 million steps (Turner et al., 2020).

Concept erasure in text-to-image diffusion models produces a different class of external side-effects. The SEE benchmark evaluates three aspects: impact on neighboring concepts, evasion of targets, and attribute leakage (Saha et al., 20 Aug 2025). It uses 5,056 prompts over 79 MS-COCO categories in 11 superclasses, with 64 attribute combinations per object, and generates 20,224 images per model (Saha et al., 20 Aug 2025). The main metrics are Target Accuracy, where lower is better, and Preserve Accuracy, where higher is better (Saha et al., 20 Aug 2025). The reported pattern is a trade-off: stronger erasure can reduce target generation but degrade preserve accuracy. For example, RECE attains target accuracy $A$1 on $A$2 and preserve accuracy $A$3 on $A$4, while UCE attains $A$5 and $A$6 respectively (Saha et al., 20 Aug 2025). Attribute leakage rises from about 35% in the unedited model to 52–61% under concept erasure techniques (Saha et al., 20 Aug 2025).

Video editing work uses the term in a still more literal sense. ROSE treats shadows, reflections, light effects, translucency, and mirror effects as object side effects that must be removed together with the object (Miao et al., 26 Aug 2025). The model uses a diffusion transformer, full-video reference input, mask augmentation, and a difference mask predictor. The ground-truth difference mask is defined by thresholding the per-pixel $A$7 difference between the original and erased videos, and training minimizes a denoising loss plus a difference-mask prediction loss (Miao et al., 26 Aug 2025). On synthetic ROSE-Bench, ROSE reports mean performance of 31.12 PSNR, 0.917 SSIM, and 0.077 LPIPS, compared with 26.50, 0.898, and 0.128 for DiffuEraser (Miao et al., 26 Aug 2025).

These results show that external side-effects in AI need not be limited to environmental disruption by agents. They also include representational spillover in edited models, collateral degradation of preserved capabilities, and failure modes that become visible only when evaluation moves beyond the immediate target prompt or mask.

5. Physical systems, therapy, and environmental spillover

In accelerator physics, the hollow electron beam in the CSRm electron cooler improves beam quality but also introduces nonlinear electromagnetic fields causing tuneshift, tunespread, and resonance driving terms (Tang et al., 2015). For $A$8 ions at 1.272 MeV/u, the hollow beam yields $A$9 and $M~A$0, while the solid beam yields $M~A$1 and $M~A$2 (Tang et al., 2015). The hollow beam produces a wider tune spread area but lower maximum tuneshift, and under the studied parameters “do not lead to serious resonances” (Tang et al., 2015). Here the external side-effect is a by-product of the cooling apparatus acting on beam dynamics beyond the intended quality improvement.

Plasmonic photothermal therapy in brain tissue offers a biomedical analogue. The therapeutic mechanism is local heating by plasmonic nanoparticles under near-infrared laser irradiation, but surrounding brain tissue can be damaged because of “absorption of laser light, thermal conductivity, nanoparticles diffusing from the tumor, or a combination hereof” (He et al., 2019). In porcine brain tissue, heating rates under NIR laser irradiation were $M~A$3 K/W for cerebrum, $M~A$4 K/W for cerebellum, and $M~A$5 K/W for brain stem, with the lower brain-stem heating attributed to higher scattering (He et al., 2019). The temperature dynamics were modeled as

$M~A$6

for heating and

$M~A$7

for cooling, with the photothermal source term

$M~A$8

and environmental heat loss $M~A$9 (He et al., 2019).

These physical examples reinforce a general point: external side-effects often arise through transport processes—field propagation, thermal conduction, diffusion, or resonance coupling—that are orthogonal to the nominal function of the intervention. This suggests that side-effect analysis in such systems is fundamentally a problem of system-environment coupling.

6. Assessment, trade-offs, and recurrent controversies

Evaluation of external side-effects typically requires measurements that compare intended performance with wider-system disturbance. In RL, SafeLife quantifies side effects as the Wasserstein-1 distance between the final pattern of protected cells and the pattern under no agent intervention (Turner et al., 2020). In concept erasure, SEE combines hierarchical prompts, CLIP zero-shot classification, and VQA-based verification to quantify target suppression, collateral damage, and attribute leakage (Saha et al., 20 Aug 2025). In ROSE, PSNR, SSIM, LPIPS, and VBench are used to compare object removal with and without correction of side effects such as shadows or reflections (Miao et al., 26 Aug 2025). In pharmacovigilance, Overlap, mention frequency over 14-day intervals, and ASE-ASE co-occurrence networks are used to validate whether social media surfaces established and novel adverse side effects (Bartal et al., 2024).

One controversy concerns whether real-world data sources can support reliable side-effect discovery. The social-media studies argue that large-scale Reddit and $\text{IO}~A$0 data can complement standard surveillance by surfacing signals earlier or in patient-centered language (Bartal et al., 2024, Sehgal et al., 12 Mar 2026, Duan et al., 6 Apr 2025). At the same time, they explicitly note limitations: voluntary and self-selected reporting, lack of causal identification, demographic skew, and possible NLP misclassification (Sehgal et al., 12 Mar 2026). The appropriate interpretation is therefore complementary rather than substitutional.

Another controversy concerns effectiveness versus side effects in regulatory technology. Age Assurance Technologies are assessed with respect to effectiveness, side effects, and acceptance, with the most important side effects involving privacy and anonymity, bias, discrimination, exclusion, and censorship (Lueks et al., 26 Mar 2026). Online architectures create tracking and profiling risks because a third-party age assurance service or identity provider learns which services users access and when (Lueks et al., 26 Mar 2026). AI age estimation can be less accurate for certain demographics, and NIST reports that even the best facial age estimators for teens have a mean absolute error of 3–5 years, with fewer than 35% of 13-year-olds correctly classified within one year of true age (Lueks et al., 26 Mar 2026). The paper therefore recommends a graduated hierarchy favoring on-device parental control and consent, device-based age assurance, and credential-based solutions with selective disclosure and unlinkability over online assurance, especially document upload to private service providers (Lueks et al., 26 Mar 2026).

A final recurrent theme is that narrow evaluation understates external side-effects. Concept erasure evaluations that check only direct target prompts miss impacts on semantic neighbors and compositional variants (Saha et al., 20 Aug 2025). Video object removal evaluations that focus on the object mask miss shadows, light, and mirror effects (Miao et al., 26 Aug 2025). Drug safety pipelines that rely only on labels and trials miss externally reported symptoms from real-world use (Bartal et al., 2024, Sehgal et al., 12 Mar 2026). This suggests that side-effect analysis is, in practice, an evaluation design problem: the wider the context represented in the benchmark, the more of the externality becomes visible.