Summary: Even if initial accuracy is high, real-world data constantly changes like a "dialect," quietly degrading the model. If left unchecked, this poses risks such as violations of international regulations like the EU AI Act or brand damage due to unintentional discrimination. This article clarifies the nature of AI "drift" through four categories and explains in an easy-to-understand way the new standard of "responsible AI operations" that both engineers and management should be aware of to maintain AI quality.%3Ch1%3ETable%20of%20contents%3C%2Fh1%3E1. What is Drift?1.1. IntroductionYou’ve spoken Japanese for years — maybe as your native language, maybe after living here long enough to negotiate with landlords and understand NHK news.Then one day, you’re in rural Tōhoku or Kagoshima. You try ordering at a soba shop or asking someone for directions. The grammar is familiar, the words are vaguely recognizable, but the meaning doesn’t land. You're speaking the “same language,” but the communication fails in a way you didn't expect.It’s not your fault. And it’s not theirs. It’s because of a gradual shift in semantics, because of time or geography.This is what AI models experience in the wild. It’s like they were trained in Tokyo Japanese — and suddenly, one region speaks Yamagata-ben.That mismatch, between what the model expects and what it encounters, is called drift.1.2. What is drift, really?In machine learning, drift refers to changes in the input data or the relationship between input and output after a model is deployed.Drift is also commonly referred to as model drift, data drift, concept drift, or even contextual drift in industry discussions. “Model drift” is often used as a catch-all phrase for performance degradation over time, while terms like “data drift” and “concept drift” describe more specific causes. “Contextual drift” is less formal in the research literature but sometimes used to describe changes in the surrounding environment that affect model performance.The coexistence of these overlapping terms often creates confusion, since different teams may use the same word to mean different things. For example, “model drift” might be logged in a monitoring dashboard, but the root cause could be data drift or concept drift.In this article, we’ll cut through that ambiguity and focus on the more precise categories — data drift, concept drift, out-of-distribution drift, and feature drift — so that QA (Quality Assurance）and engineering teams can diagnose and address the right problems systematically.2. What is This Article For?2.1 Objectives- Explain different types of drift in AI systems- Show why monitoring drift is critical, especially in high-risk applications- Lay the groundwork for strategies to detect, monitor, and mitigate drift (covered in the next article in more detail)2.2 Who Can Benefit From This Article- Software engineers deploying AI in production- QA(Quality Assurance) or QC(Quality Control) professionals ensuring reliability- ML engineers building or maintaining models- Risk & compliance officers working under the EU AI Act or similar regulations3. What Are the Types of Drift?Before knowing how to avoid drift, we need to clarify first the types of drift. Each type corresponds to a distinct form of mismatch between the data distribution seen during training and the data encountered during deployment. These mismatches require different detection methods and mitigation strategies.3.1 Data Drift (a.k.a. Covariate Drift)Data drift occurs when the statistical distribution of the input data shifts over time, even though the model’s task and the label space remain the same.Examples:- Handwriting Recognition: Government OCR systems trained on younger handwriting fail to read more wobbly characters as Japan’s elderly population increases from submissions.- Road Sensor Data: In autonomous driving, an AV trained in Tokyo might misinterpret rural road markings or faded signage in other prefectures.- Camera Conditions: A speed detection model trained on daytime RGB images may perform poorly at night, when lighting conditions differ. Why it matters:If your model assumes that inputs remain static but reality shifts, performance will quietly degrade. Retraining without understanding the shift can lead to garbage-in-garbage-out loops.3.2 Concept DriftConcept drift occurs when the relationship between input and output changes over time. The data distribution might look familiar (unlike in data drift), but the meaning behind it — the way the model should interpret it — has shifted.Unlike data drift, which is about how the inputs change, concept drift is about how the same inputs now lead to different outputs. This means the data hasn’t changed but the interpretation of it has.Examples:- Smart Meter Anomaly Detection: A model flags households for unusually high or low energy use to detect theft or faulty meters. But with solar panels and EV chargers, households now have irregular usage — spikes at night for charging, zero midday due to solar. Daily kWh totals may look similar, but their interpretation has changed. Models miss fraud or overflag normal homes.- Classroom Attention Detection: An AI model equates “still, eye contact, facing forward” with focus. But for many students with ADHD, fidgeting helps them concentrate. The same behaviors, once treated as distraction, now reflect deeper engagement.- Sentiment Analysis & Slang: Older models might classify phrases like “It’s insane” as negative, since “insane” historically carried a mental health stigma. But in modern slang, “That game was insane” means exciting or excellent. Similarly in Japanese, やばい (yabai) once meant “dangerous” or “risky,” but in everyday speech it can now mean “amazing” or “cool.” Same word, completely different polarity.Why it matters:Concept drift is insidious because inputs look normal, but the labels and outcomes evolve. Your system silently mispredicts while everything appears fine.3.3 Out-of-Distribution (OOD) DriftOOD drift describes cases where inputs fall entirely outside the model's training distribution. They’re data points the model was never designed to understand or never encountered before.Examples:- Natural Disasters: A landslide destroys a road and the AV navigation model sees terrain it's never been trained on, leading to misclassification.- Invasive Weed Species in Agriculture AI: An agricultural vision system trained only on known crops and weeds mislabels a newly introduced invasive plant. But the system ignores it entirely or misclassifies it as a harmless species.- Healthcare System Shock: A new disease emerges (e.g., during a pandemic), and symptom checkers or triage models receive inputs that don’t match any known diagnostic paths.Why it matters:OOD events often occur during shocks or rare events, when systems are most vulnerable. Because the model has zero exposure to the input type, predictions become erratic or overconfidently wrong.3.4 Feature Drift (a Subtype of Data Drift, a.k.a. Univariate Drift)Feature drift is a more focused case of data drift. It happens when the statistical distribution of one or a few features changes significantly, even if the overall dataset looks stable. While data drift highlights global changes, feature drift drills down to the level of individual attributes.Examples:- Job Title Drift (Resume Screening): Models trained on titles like “Software Engineer” or “Data Scientist” struggle when new roles such as “AI Product Lead” or “Prompt Engineer” become common. A single feature changes, and the model underperforms even though the rest of the dataset seems stable.- Age Distribution Drift: An applicant pool suddenly skews younger (due to more graduates entering the market) or older (due to mid-career shifts). The age feature drifts, and if age is correlated with outcomes, it can bias hiring decisions.- Migration Flux Drift: Workforce demographics change as more foreign applicants apply for jobs. If the nationality or language background feature shifts, models may unfairly penalize non-domestic applicants — not because of qualifications, but due to historical bias encoded in training data.Why it matters:You may think that feature drift can be ignored and that only global data drift deserves attention, since the latter often shows a bigger impact on performance. However, overlooking feature drift hides critical risks:- Overall performance may still look acceptable, while bias builds up silently in a single attribute.- Sensitive features like age, nationality, or job title may drift in ways that amplify discrimination, even if other features are stable.- From a compliance standpoint, regulators don’t just want to know that the dataset shifted; they want to know which features shifted and why that matters for fairness and accountability.Monitoring only at the global level risks missing exactly the kinds of harms the EU AI Act and similar frameworks are designed to prevent. Feature drift monitoring is therefore essential for pinpointing root causes and ensuring systems remain both effective and equitable.3.5. Summary TableType of DriftDefinitionExamplesData Drift (Covariate Drift)Input data distribution changes, but the relationship to labels stays the same.- Handwriting Recognition: OCR trained on younger handwriting fails as elderly submissions increase.- Road Sensor Data: AV trained in Tokyo misreads rural markings/faded signs.- Camera Conditions: Speed detection trained on daytime images fails at night.Concept DriftThe relationship between inputs and outputs changes, even if inputs look similar.- Smart meter anomaly detection: abnormal household electricity use shifts from late-night gaming to daytime EV charging.- Sentiment analysis: phrases like “It’s insane” or やばい (yabai) flip from negative to positive meanings.- Classroom focus detection: students with ADHD show more movement, which is not necessarily a sign of distraction.Out-of-Distribution (OOD) DriftModel encounters inputs completely outside the training distribution.- Healthcare Shocks: New disease emerges, symptom checkers fail on unseen patterns.- Agriculture AI mislabels newly introduced invasive plants as known crops/weeds.- Autonomous driving misclassifies terrain after a landslide.Feature Drift (Subtype of Data Drift)Distribution of specific features shifts, even if the global dataset looks stable.- Resume screening: new job titles (e.g., AI Product Lead) not recognized by older models.- Age distribution: applicant pool shifts younger or older, biasing outcomes.- Migration flux: increase in foreign applicants triggers bias in models not designed for demographic change.In this article, we examined the mechanisms behind model drift and their characteristics. In our next post, we will discuss how these bottlenecks pose regulatory and operational risks in high-risk AI applications, and why QA and explainability are so crucial.References [1] EU AI Act, Article 10: Data and Data Governance (Link)[2] EU AI Act, Article 16: Obligations of Providers of High-Risk AI Systems (Link)[3] EU AI Act, Article 17: Quality Management System (Link)[4] EU AI Act, Article 72: Post-Market Monitoring by Providers and Post-Market Monitoring Plan for High-Risk AI Systems (Link)[5] EU AI Act, Article 99: Penalties (Link)[6] Labor Standards Act (Act No. 49 of 1947), Article 3 (Link)