Corey B

Lab / Carbon / Methodology

Where these numbers come from

Every slider in the Lab traces back to a primary source. Some of those sources are rock-solid. Some span three orders of magnitude and the honest answer is “nobody really knows yet.” This page lays out which is which — so you can decide how much to trust the number staring back at you.

Travel

Gasoline cars use the EPA fleet-average figure of 404 gCO₂e per mile — a 22.8 mpg combined average with CH₄ and N₂O adjustments baked in. Real vehicles span roughly 0.22 kg/mi (compact hybrid) to 0.55 kg/mi (full-size pickup); the Lab uses the fleet average unless you tell it otherwise.

Electric vehicles are entirely grid-dependent. The Lab assumes 0.30 kWh/mi (EPA Fuel Economy average BEV) and multiplies by your eGRID subregion intensity. The spread is real: a PNW-grid EV emits ~0.03 kg/mi, a coal-heavy grid EV can exceed 0.25 kg/mi — nearly an order of magnitude apart.

Transit numbers come from the EPA Emission Factors Hub (bus 0.066, subway 0.096, intercity rail 0.133 kgCO₂e per passenger-mile). These are fleet averages; a packed peak-hour subway is far better per passenger than an empty late-night one, but we don’t have your specific train.

Flights include radiative forcing by default. At cruise altitude, NOₓ emissions and contrails trigger warming effects roughly as large as the CO₂ itself. Ignoring that doubling is scientifically indefensible, even though what you see on the slider isn’t strictly “CO₂.” We use DEFRA 2024 with-RF factors (short-haul 0.246, long-haul 0.311 kgCO₂e/pax-mile, economy). Business class is roughly 2× and first ~4× due to seat footprint — v1 only models economy.

Home

Electricity is the slider with the most regional variance in the whole Lab. The US average is 0.385 kgCO₂e/kWh (EPA eGRID 2022), but subregions range from ~0.18 (NYC, New England) through ~0.22 (California) to ~0.62 (lower Midwest / Ohio) and ~0.72 (Oahu). That’s a 4–5× spread. The Lab asks you to pick an eGRID subregion rather than hiding this behind a national average, because a national average quietly lies to roughly everyone.

Natural gas is a clean 5.30 kgCO₂e/therm (EPA Equivalencies). Heating oil is 10.28 kg/gallon (EIA). Propane is 5.76 kg/gallon (EPA Emission Factors Hub 2025). These three are well-measured and have narrow uncertainty bands.

Home size is a rough scalar on electricity + heating draw, grounded in EIA RECS 2020 end-use intensity by floor area (roughly 0.6× under 800 sqft, 1.0× 800–1800, 1.4× 1800–3000, 1.8× above). A proper model would split heating, cooling, and water heating separately and factor climate zone. We don’t; a v2 might.

Food

All food numbers come from Poore & Nemecek (2018, Science) — a meta-analysis of 38,700 farms across 119 countries and 40 products, cradle-to-retail. It’s still the best single source on food LCA, and it’s what Our World in Data uses.

The headlines are stark. Beef from dedicated beef herds lands at ~60 kgCO₂e/kg — roughly 60× peas. Beef from dairy herds is ~35% of that because the impact is co-allocated with milk, so the Lab uses a blended ~35 kg/kg by default. Pork is ~7, chicken ~6, farmed fish ~5, eggs ~4.5, milk ~3, cheese ~21. Most plant staples are under 5; vegetables average 0.5.

Servings-per-week entry uses USDA-typical portion sizes (113 g ≈ ¼ lb for meat, 244 g for a cup of milk, 30 g for an ounce of cheese) so you don’t have to weigh anything to get a sane annual total.

Stuff

This is the roughest category in the Lab and it’s worth saying so out loud. Embodied emissions for physical goods vary enormously with fabric, geography, manufacturing process, and use-phase behavior (how many times you wash the shirt matters a lot).

Per-item numbers: a cotton t-shirt averages ~15 kgCO₂e (Carbon Trust), polyester ~5.5, a pair of jeans ~25 (Levi’s LCA + UCL 2024 meta-analysis). Smartphones land at ~55 kg embodied, with manufacturing being 85–95% of lifetime footprint — your refresh cadence dominates the number. Laptops are ~200 kg embodied (Apple and Dell product environmental reports).

“Consumer goods by dollars spent” uses an economic input-output LCA proxy of 0.4 kgCO₂e per US dollar of non-food retail spending. This is useful at the how much do I spend on stuff per year scale and essentially useless at the per-purchase level. We present it as order-of-magnitude.

Digital & AI

Per-query AI emissions are genuinely unsettled. Public estimates span roughly three orders of magnitude — from Google’s 2025 Gemini disclosure of 0.03 gCO₂e per median text prompt, up through ~3 g for a frontier chat on a US-average grid, up to ~30 g for a long reasoning response on a dirty grid. The Lab surfaces a range for every AI slider rather than pretending we have a point estimate. Showing the range is the honest move.

Five variables drive the spread, roughly in decreasing order of impact: grid carbon intensity at the serving data center (~4–5×), model size and architecture (10–100× between a small open model and a frontier chat model), hardware generation (H100 ≫ A100 ≫ older), output length (reasoning models emit 10–30× more tokens), and serving-efficiency details like quantization and batching (another 2–5×).

Image generation lands around 5 gCO₂e for an SDXL-class 512² generation on the US-average grid, with a ~10× spread across models on the HF AI Energy Score leaderboard. Video generation is super-linear in duration: a 10-second clip is not 10× a 1-second clip — it’s closer to 15–20× because denoising scales with frame count times resolution times steps.

What about training? Frontier-model training runs are in the 500–6000 tCO₂e range. Amortized across 10¹ⁱ–10¹² lifetime queries that’s single-digit milligrams per query — roughly 1–2 orders of magnitude less than inference. The calculator counts inference, not training, because counting both would be double- charging for a cost that’s already priced into the per-query number at scale.

Mitigations & offsets

This is where the honest-tone rule matters most. A calculator that hands you a single “you offset X tonnes” number without the caveats below is helping you lie to yourself.

Solar, heat pumps, EVs. All three displace grid electricity or fossil fuels, so the benefit keys off the same eGRID subregion you picked in Home. A typical 8 kW residential solar install generates ~11,000 kWh/year (NREL PVWatts) and avoids ~4.3 tCO₂e/year on the US-average grid — ~2 t on a clean grid, ~8 t on a coal-heavy one. Heat pumps replacing gas average ~2.3 t/year avoided but can be near break-even in a mild climate on a coal grid (and over 3 t in a cold climate on a clean grid). EVs replacing a 22.8 mpg ICE average ~3.5 t/year avoided; embodied battery emissions (~6–10 t) amortize over vehicle life and typically pay back in 1–2 years of driving.

Dietary shifts come from applying Poore & Nemecek intensities to USDA average US intakes. Beef → chicken is ~1.4 t/year (the biggest single food lever), omnivore → vegetarian adds another ~0.8 t, and vegetarian → vegan another ~0.4 t. Scarborough et al. 2023 (Nature Food) puts the full high-meat-to-vegan delta at ~1.5 t/year, which is consistent with summing these steps. Note the vegetarian step is smaller than people expect because cheese is ~21 kgCO₂e/kg and Americans eat ~17 kg/year of it.

Tree planting has a 20-year lag. A sapling sequesters essentially zero CO₂ in its first year. A temperate hardwood doesn’t hit meaningful sequestration (~20 kg/year) until year 10–15. Averaged across 20 years and discounting for the ~25% of planted trees that don’t survive, a typical planted tree locks away about 0.2 tonnes over 20 years. The slider uses that figure because that’s the honest answer to “plant a tree today, how much does it help?” Trees also burn, get cut down, and release their carbon on death — permanence is not guaranteed.

Offsets are three different things with three different credibility levels. The Lab multiplies any offset spend by a credibility factor before subtracting:

  • Verified engineered removal (DAC, enhanced weathering) at $400–1000/tonne — credibility 0.95. Physically removes CO₂, stores it in geology for ~10,000 years. The real thing, and priced that way.
  • Nature-based, independently verified (Verra VCS, Gold Standard, ACR) at $15–60/tonne — credibility 0.35. We hard-haircut this tier because the evidence demands it: the Guardian / Die Zeit / SourceMaterial 2023 investigation found >90% of Verra REDD+ credits didn’t represent real reductions, and West et al. 2023 (Science) independently put the figure at ~94% for the largest projects. 0.35 is already generous; a strict read puts it closer to 0.10.
  • Unverified / cheap retail (airline checkout, $3–10/tonne) — credibility 0.05. Real marginal abatement costs more than this; credits at $5/tonne are almost always non-additional, double-counted, or fabricated. Treat as effectively zero.

Bottom line: only engineered removals meaningfully draw down CO₂ that’s already in the atmosphere. Reducing your own emissions through the other five categories is always worth more per dollar than nature-based offsets, and orders of magnitude more than cheap retail ones.

What this calculator doesn't do

An honest methodology page has to include its own shortcomings. v1 scope is deliberately narrow. Here’s what it doesn’t cover:

  • US-only. eGRID subregions, EPA transit factors, EIA fuel factors, and USDA food intakes are all US. Non-US users will get a rough directional answer, not a calibrated one.
  • Economy class flights only. Business is ~2×, first ~4× due to seat footprint. If you flew business, your travel number is lowballed.
  • No household-level modeling. Everything is per-person. A four-person household sharing a furnace double-counts heating if everyone runs the calculator independently.
  • Stuff is an EIO proxy. Accurate for an annual spending bucket, useless for any individual purchase. Don’t use it to A/B test two sweaters.
  • Home size is a flat scalar. It doesn’t split heating vs cooling vs water-heating, and it ignores climate zone. HDD-weighted modeling is a v2 problem.
  • AI emissions are a range, not a number. You will see a low / central / high band. Picking the middle and treating it as truth is exactly the thing this page is trying to avoid.
  • No Scope 3 supply chain for services. Streaming, cloud storage, and non-AI digital usage use back-of-envelope factors. The resolution isn’t there to make fine-grained claims.
  • Not peer-reviewed. This is one person’s Earth Day project. The numbers trace back to primary sources linked above; the synthesis is mine. Audit it against the citations and tell me if something’s wrong.