Most of eastern North America receives elevated levels of atmospheric deposition of sulfur (S) that result from anthropogenic SO₂ emissions from fossil fuel combustion. Atmospheric S deposition has acidified sensitive terrestrial and aquatic ecosystems in this region; however, deposition has been declining since the 1970s, resulting in some recovery in previously acidified aquatic ecosystems. Accurate watershed S mass balances help to evaluate the extent to which atmospheric S deposition is retained within ecosystems, and whether internal cycling sources and biogeochemical processes may be affecting the rate of recovery from decreasing S atmospheric loads. This study evaluated S mass balances for 15 sites with watersheds in southeastern Canada and northeastern US for the period 1985 to 2002. These 15 sites included nine in Canada (Turkey Lakes, ON; Harp Lake, ON; Plastic Lake, ON; Hermine, QC; Lake Laflamme, QC; Lake Clair, QC; Lake Tirasse, QC; Mersey, NS; Moosepit, NS) and six in the US (Arbutus Lake, NY; Biscuit Brook, NY; Sleepers River, VT; Hubbard Brook Experimental Forest, NH; Cone Pond, NH; Bear Brook Watershed, ME). Annual S wet deposition inputs were derived from measured bulk or wet-only deposition and stream export was obtained by combining drainage water fluxes with SO₄ ²⁻ concentrations. Dry deposition has the greatest uncertainty of any of the mass flux calculations necessary to develop accurate watershed balances, and here we developed a new method to calculate this quantity. We utilized historical information from both the US National Emissions Inventory and the US (CASTNET) and the Canadian (CAPMoN) dry deposition networks to develop a formulation that predicted SO₂ concentrations as a function of SO₂ emissions, latitude and longitude. The SO₂ concentrations were used to predict dry deposition using relationships between concentrations and deposition flux derived from the CASTNET or CAPMoN networks. For the year 2002, we compared the SO₂ concentrations and deposition predictions with the predictions of two continental-scale air quality models, the Community Multiscale Air Quality (CMAQ) model and A Unified Regional Air-quality Modeling System (AURAMS) that utilize complete inventories of emissions and chemical budgets. The results of this comparison indicated that the predictive relationship provides an accurate representation of SO₂ concentrations and S deposition for the region that is generally consistent with these models, and thus provides confidence that our approach could be used to develop accurate watershed S budgets for these 15 sites. Most watersheds showed large net losses of SO₄ ²⁻ on an annual basis, and the watershed mass balances were grouped into five categories based on the relative value of mean annual net losses or net gains. The net annual fluxes of SO₄ ²⁻ showed a strong relationship with hydrology; the largest net annual negative fluxes were associated with years of greatest precipitation amount and highest discharge. The important role of catchment hydrology on S budgets suggests implications for future predicted climate change as it affects patterns of precipitation and drought. The sensitivity of S budgets is likely to be greatest in watersheds with the greatest wetland area, which are particularly sensitive to drying and wetting cycles. A small number of the watersheds in this analysis were shown to have substantial S sources from mineral weathering, but most showed evidence of an internal source of SO₄ ²⁻, which is likely from the mineralization of organic S stored from decades of increased S deposition. Mobilization of this internal S appears to contribute about 1–6 kg S ha⁻¹ year⁻¹ to stream fluxes at these sites and is affecting the rate and extent of recovery from acidification as S deposition rates have declined in recent years. This internal S source should be considered when developing critical deposition loads that will promote ecosystem recovery from acidification and the depletion of nutrient cations in the northeastern US and southeastern Canada.